std/sys/thread/
unix.rs

1#[cfg(not(any(
2    target_env = "newlib",
3    target_os = "l4re",
4    target_os = "emscripten",
5    target_os = "redox",
6    target_os = "hurd",
7    target_os = "aix",
8    target_os = "wasi",
9)))]
10use crate::ffi::CStr;
11use crate::mem::{self, DropGuard, ManuallyDrop};
12use crate::num::NonZero;
13#[cfg(all(target_os = "linux", target_env = "gnu"))]
14use crate::sys::weak::dlsym;
15#[cfg(any(target_os = "solaris", target_os = "illumos", target_os = "nto",))]
16use crate::sys::weak::weak;
17use crate::thread::ThreadInit;
18use crate::time::Duration;
19use crate::{cmp, io, ptr, sys};
20#[cfg(not(any(
21    target_os = "l4re",
22    target_os = "vxworks",
23    target_os = "espidf",
24    target_os = "nuttx"
25)))]
26pub const DEFAULT_MIN_STACK_SIZE: usize = 2 * 1024 * 1024;
27#[cfg(target_os = "l4re")]
28pub const DEFAULT_MIN_STACK_SIZE: usize = 1024 * 1024;
29#[cfg(target_os = "vxworks")]
30pub const DEFAULT_MIN_STACK_SIZE: usize = 256 * 1024;
31#[cfg(any(target_os = "espidf", target_os = "nuttx"))]
32pub const DEFAULT_MIN_STACK_SIZE: usize = 0; // 0 indicates that the stack size configured in the ESP-IDF/NuttX menuconfig system should be used
33
34pub struct Thread {
35    id: libc::pthread_t,
36}
37
38// Some platforms may have pthread_t as a pointer in which case we still want
39// a thread to be Send/Sync
40unsafe impl Send for Thread {}
41unsafe impl Sync for Thread {}
42
43impl Thread {
44    // unsafe: see thread::Builder::spawn_unchecked for safety requirements
45    #[cfg_attr(miri, track_caller)] // even without panics, this helps for Miri backtraces
46    pub unsafe fn new(stack: usize, init: Box<ThreadInit>) -> io::Result<Thread> {
47        // FIXME: remove this block once wasi-sdk is updated with the fix from
48        // https://github.com/WebAssembly/wasi-libc/pull/716
49        // WASI does not support threading via pthreads. While wasi-libc provides
50        // pthread stubs, pthread_create returns EAGAIN, which causes confusing
51        // errors. We return UNSUPPORTED_PLATFORM directly instead.
52
53        // NOTE: exempt `wasm32-wasip1-threads` from this check as `emnapi` has a working pthread
54        // implementation. See <https://github.com/rust-lang/rust/issues/153475>.
55        if cfg!(all(target_os = "wasi", not(all(target_env = "p1", target_feature = "atomics")))) {
56            return Err(io::Error::UNSUPPORTED_PLATFORM);
57        }
58
59        let data = init;
60        let mut attr: mem::MaybeUninit<libc::pthread_attr_t> = mem::MaybeUninit::uninit();
61        assert_eq!(libc::pthread_attr_init(attr.as_mut_ptr()), 0);
62        let mut attr = DropGuard::new(&mut attr, |attr| {
63            assert_eq!(libc::pthread_attr_destroy(attr.as_mut_ptr()), 0)
64        });
65
66        #[cfg(any(target_os = "espidf", target_os = "nuttx"))]
67        if stack > 0 {
68            // Only set the stack if a non-zero value is passed
69            // 0 is used as an indication that the default stack size configured in the ESP-IDF/NuttX menuconfig system should be used
70            assert_eq!(
71                libc::pthread_attr_setstacksize(
72                    attr.as_mut_ptr(),
73                    cmp::max(stack, min_stack_size(attr.as_ptr()))
74                ),
75                0
76            );
77        }
78
79        #[cfg(not(any(target_os = "espidf", target_os = "nuttx")))]
80        {
81            let stack_size = cmp::max(stack, min_stack_size(attr.as_ptr()));
82
83            match libc::pthread_attr_setstacksize(attr.as_mut_ptr(), stack_size) {
84                0 => {}
85                n => {
86                    assert_eq!(n, libc::EINVAL);
87                    // EINVAL means |stack_size| is either too small or not a
88                    // multiple of the system page size. Because it's definitely
89                    // >= PTHREAD_STACK_MIN, it must be an alignment issue.
90                    // Round up to the nearest page and try again.
91                    let page_size = sys::os::page_size();
92                    let stack_size =
93                        (stack_size + page_size - 1) & (-(page_size as isize - 1) as usize - 1);
94
95                    // Some libc implementations, e.g. musl, place an upper bound
96                    // on the stack size, in which case we can only gracefully return
97                    // an error here.
98                    if libc::pthread_attr_setstacksize(attr.as_mut_ptr(), stack_size) != 0 {
99                        return Err(io::const_error!(
100                            io::ErrorKind::InvalidInput,
101                            "invalid stack size"
102                        ));
103                    }
104                }
105            };
106        }
107
108        let data = Box::into_raw(data);
109        let mut native: libc::pthread_t = mem::zeroed();
110        let ret = libc::pthread_create(&mut native, attr.as_ptr(), thread_start, data as *mut _);
111        return if ret == 0 {
112            Ok(Thread { id: native })
113        } else {
114            // The thread failed to start and as a result `data` was not consumed.
115            // Therefore, it is safe to reconstruct the box so that it gets deallocated.
116            drop(Box::from_raw(data));
117            Err(io::Error::from_raw_os_error(ret))
118        };
119
120        extern "C" fn thread_start(data: *mut libc::c_void) -> *mut libc::c_void {
121            unsafe {
122                // SAFETY: we are simply recreating the box that was leaked earlier.
123                let init = Box::from_raw(data as *mut ThreadInit);
124                let rust_start = init.init();
125
126                // Now that the thread information is set, set up our stack
127                // overflow handler.
128                let _handler = sys::stack_overflow::Handler::new();
129
130                rust_start();
131            }
132            ptr::null_mut()
133        }
134    }
135
136    pub fn join(self) {
137        let id = self.into_id();
138        let ret = unsafe { libc::pthread_join(id, ptr::null_mut()) };
139        assert!(ret == 0, "failed to join thread: {}", io::Error::from_raw_os_error(ret));
140    }
141
142    #[cfg(not(target_os = "wasi"))]
143    pub fn id(&self) -> libc::pthread_t {
144        self.id
145    }
146
147    pub fn into_id(self) -> libc::pthread_t {
148        ManuallyDrop::new(self).id
149    }
150}
151
152impl Drop for Thread {
153    fn drop(&mut self) {
154        let ret = unsafe { libc::pthread_detach(self.id) };
155        debug_assert_eq!(ret, 0);
156    }
157}
158
159pub fn available_parallelism() -> io::Result<NonZero<usize>> {
160    cfg_select! {
161        any(
162            target_os = "android",
163            target_os = "emscripten",
164            target_os = "fuchsia",
165            target_os = "hurd",
166            target_os = "linux",
167            target_os = "aix",
168            target_vendor = "apple",
169            target_os = "cygwin",
170        ) => {
171            #[allow(unused_assignments)]
172            #[allow(unused_mut)]
173            let mut quota = usize::MAX;
174
175            #[cfg(any(target_os = "android", target_os = "linux"))]
176            {
177                quota = cgroups::quota().max(1);
178                let mut set: libc::cpu_set_t = unsafe { mem::zeroed() };
179                unsafe {
180                    if libc::sched_getaffinity(0, size_of::<libc::cpu_set_t>(), &mut set) == 0 {
181                        let count = libc::CPU_COUNT(&set) as usize;
182                        let count = count.min(quota);
183
184                        // According to sched_getaffinity's API it should always be non-zero, but
185                        // some old MIPS kernels were buggy and zero-initialized the mask if
186                        // none was explicitly set.
187                        // In that case we use the sysconf fallback.
188                        if let Some(count) = NonZero::new(count) {
189                            return Ok(count)
190                        }
191                    }
192                }
193            }
194            match unsafe { libc::sysconf(libc::_SC_NPROCESSORS_ONLN) } {
195                -1 => Err(io::Error::last_os_error()),
196                0 => Err(io::Error::UNKNOWN_THREAD_COUNT),
197                cpus => {
198                    let count = cpus as usize;
199                    // Cover the unusual situation where we were able to get the quota but not the affinity mask
200                    let count = count.min(quota);
201                    Ok(unsafe { NonZero::new_unchecked(count) })
202                }
203            }
204        }
205        any(
206           target_os = "freebsd",
207           target_os = "dragonfly",
208           target_os = "openbsd",
209           target_os = "netbsd",
210        ) => {
211            use crate::ptr;
212
213            #[cfg(target_os = "freebsd")]
214            {
215                let mut set: libc::cpuset_t = unsafe { mem::zeroed() };
216                unsafe {
217                    if libc::cpuset_getaffinity(
218                        libc::CPU_LEVEL_WHICH,
219                        libc::CPU_WHICH_PID,
220                        -1,
221                        size_of::<libc::cpuset_t>(),
222                        &mut set,
223                    ) == 0 {
224                        let count = libc::CPU_COUNT(&set) as usize;
225                        if count > 0 {
226                            return Ok(NonZero::new_unchecked(count));
227                        }
228                    }
229                }
230            }
231
232            #[cfg(target_os = "netbsd")]
233            {
234                unsafe {
235                    let set = libc::_cpuset_create();
236                    if !set.is_null() {
237                        let mut count: usize = 0;
238                        if libc::pthread_getaffinity_np(libc::pthread_self(), libc::_cpuset_size(set), set) == 0 {
239                            for i in 0..libc::cpuid_t::MAX {
240                                match libc::_cpuset_isset(i, set) {
241                                    -1 => break,
242                                    0 => continue,
243                                    _ => count = count + 1,
244                                }
245                            }
246                        }
247                        libc::_cpuset_destroy(set);
248                        if let Some(count) = NonZero::new(count) {
249                            return Ok(count);
250                        }
251                    }
252                }
253            }
254
255            let mut cpus: libc::c_uint = 0;
256            let mut cpus_size = size_of_val(&cpus);
257
258            unsafe {
259                cpus = libc::sysconf(libc::_SC_NPROCESSORS_ONLN) as libc::c_uint;
260            }
261
262            // Fallback approach in case of errors or no hardware threads.
263            if cpus < 1 {
264                let mut mib = [libc::CTL_HW, libc::HW_NCPU, 0, 0];
265                let res = unsafe {
266                    libc::sysctl(
267                        mib.as_mut_ptr(),
268                        2,
269                        (&raw mut cpus) as *mut _,
270                        (&raw mut cpus_size) as *mut _,
271                        ptr::null_mut(),
272                        0,
273                    )
274                };
275
276                // Handle errors if any.
277                if res == -1 {
278                    return Err(io::Error::last_os_error());
279                } else if cpus == 0 {
280                    return Err(io::Error::UNKNOWN_THREAD_COUNT);
281                }
282            }
283
284            Ok(unsafe { NonZero::new_unchecked(cpus as usize) })
285        }
286        target_os = "nto" => {
287            unsafe {
288                use libc::_syspage_ptr;
289                if _syspage_ptr.is_null() {
290                    Err(io::const_error!(io::ErrorKind::NotFound, "no syspage available"))
291                } else {
292                    let cpus = (*_syspage_ptr).num_cpu;
293                    NonZero::new(cpus as usize)
294                        .ok_or(io::Error::UNKNOWN_THREAD_COUNT)
295                }
296            }
297        }
298        any(target_os = "solaris", target_os = "illumos") => {
299            let mut cpus = 0u32;
300            if unsafe { libc::pset_info(libc::PS_MYID, core::ptr::null_mut(), &mut cpus, core::ptr::null_mut()) } != 0 {
301                return Err(io::Error::UNKNOWN_THREAD_COUNT);
302            }
303            Ok(unsafe { NonZero::new_unchecked(cpus as usize) })
304        }
305        target_os = "haiku" => {
306            // system_info cpu_count field gets the static data set at boot time with `smp_set_num_cpus`
307            // `get_system_info` calls then `smp_get_num_cpus`
308            unsafe {
309                let mut sinfo: libc::system_info = crate::mem::zeroed();
310                let res = libc::get_system_info(&mut sinfo);
311
312                if res != libc::B_OK {
313                    return Err(io::Error::UNKNOWN_THREAD_COUNT);
314                }
315
316                Ok(NonZero::new_unchecked(sinfo.cpu_count as usize))
317            }
318        }
319        target_os = "vxworks" => {
320            // Note: there is also `vxCpuConfiguredGet`, closer to _SC_NPROCESSORS_CONF
321            // expectations than the actual cores availability.
322
323            // SAFETY: `vxCpuEnabledGet` always fetches a mask with at least one bit set
324            unsafe{
325                let set = libc::vxCpuEnabledGet();
326                Ok(NonZero::new_unchecked(set.count_ones() as usize))
327            }
328        }
329        _ => {
330            // FIXME: implement on Redox, l4re
331            Err(io::const_error!(io::ErrorKind::Unsupported, "getting the number of hardware threads is not supported on the target platform"))
332        }
333    }
334}
335
336pub fn current_os_id() -> Option<u64> {
337    // Most Unix platforms have a way to query an integer ID of the current thread, all with
338    // slightly different spellings.
339    //
340    // The OS thread ID is used rather than `pthread_self` so as to match what will be displayed
341    // for process inspection (debuggers, trace, `top`, etc.).
342    cfg_select! {
343        // Most platforms have a function returning a `pid_t` or int, which is an `i32`.
344        any(target_os = "android", target_os = "linux") => {
345            use crate::sys::pal::weak::syscall;
346
347            // `libc::gettid` is only available on glibc 2.30+, but the syscall is available
348            // since Linux 2.4.11.
349            syscall!(fn gettid() -> libc::pid_t;);
350
351            // SAFETY: FFI call with no preconditions.
352            let id: libc::pid_t = unsafe { gettid() };
353            Some(id as u64)
354        }
355        target_os = "nto" => {
356            // SAFETY: FFI call with no preconditions.
357            let id: libc::pid_t = unsafe { libc::gettid() };
358            Some(id as u64)
359        }
360        target_os = "openbsd" => {
361            // SAFETY: FFI call with no preconditions.
362            let id: libc::pid_t = unsafe { libc::getthrid() };
363            Some(id as u64)
364        }
365        target_os = "freebsd" => {
366            // SAFETY: FFI call with no preconditions.
367            let id: libc::c_int = unsafe { libc::pthread_getthreadid_np() };
368            Some(id as u64)
369        }
370        target_os = "netbsd" => {
371            // SAFETY: FFI call with no preconditions.
372            let id: libc::lwpid_t = unsafe { libc::_lwp_self() };
373            Some(id as u64)
374        }
375        any(target_os = "illumos", target_os = "solaris") => {
376            // On Illumos and Solaris, the `pthread_t` is the same as the OS thread ID.
377            // SAFETY: FFI call with no preconditions.
378            let id: libc::pthread_t = unsafe { libc::pthread_self() };
379            Some(id as u64)
380        }
381        target_vendor = "apple" => {
382            // Apple allows querying arbitrary thread IDs, `thread=NULL` queries the current thread.
383            let mut id = 0u64;
384            // SAFETY: `thread_id` is a valid pointer, no other preconditions.
385            let status: libc::c_int = unsafe { libc::pthread_threadid_np(0, &mut id) };
386            if status == 0 {
387                Some(id)
388            } else {
389                None
390            }
391        }
392        // Other platforms don't have an OS thread ID or don't have a way to access it.
393        _ => None,
394    }
395}
396
397#[cfg(any(
398    target_os = "linux",
399    target_os = "nto",
400    target_os = "solaris",
401    target_os = "illumos",
402    target_os = "vxworks",
403    target_os = "cygwin",
404    target_vendor = "apple",
405))]
406fn truncate_cstr<const MAX_WITH_NUL: usize>(cstr: &CStr) -> [libc::c_char; MAX_WITH_NUL] {
407    let mut result = [0; MAX_WITH_NUL];
408    for (src, dst) in cstr.to_bytes().iter().zip(&mut result[..MAX_WITH_NUL - 1]) {
409        *dst = *src as libc::c_char;
410    }
411    result
412}
413
414#[cfg(target_os = "android")]
415pub fn set_name(name: &CStr) {
416    const PR_SET_NAME: libc::c_int = 15;
417    unsafe {
418        let res = libc::prctl(
419            PR_SET_NAME,
420            name.as_ptr(),
421            0 as libc::c_ulong,
422            0 as libc::c_ulong,
423            0 as libc::c_ulong,
424        );
425        // We have no good way of propagating errors here, but in debug-builds let's check that this actually worked.
426        debug_assert_eq!(res, 0);
427    }
428}
429
430#[cfg(any(
431    target_os = "linux",
432    target_os = "freebsd",
433    target_os = "dragonfly",
434    target_os = "nuttx",
435    target_os = "cygwin"
436))]
437pub fn set_name(name: &CStr) {
438    unsafe {
439        cfg_select! {
440            any(target_os = "linux", target_os = "cygwin") => {
441                // Linux and Cygwin limits the allowed length of the name.
442                const TASK_COMM_LEN: usize = 16;
443                let name = truncate_cstr::<{ TASK_COMM_LEN }>(name);
444            }
445            _ => {
446                // FreeBSD, DragonFly BSD and NuttX do not enforce length limits.
447            }
448        };
449        // Available since glibc 2.12, musl 1.1.16, and uClibc 1.0.20 for Linux,
450        // FreeBSD 12.2 and 13.0, and DragonFly BSD 6.0.
451        let res = libc::pthread_setname_np(libc::pthread_self(), name.as_ptr());
452        // We have no good way of propagating errors here, but in debug-builds let's check that this actually worked.
453        debug_assert_eq!(res, 0);
454    }
455}
456
457#[cfg(target_os = "openbsd")]
458pub fn set_name(name: &CStr) {
459    unsafe {
460        libc::pthread_set_name_np(libc::pthread_self(), name.as_ptr());
461    }
462}
463
464#[cfg(target_vendor = "apple")]
465pub fn set_name(name: &CStr) {
466    unsafe {
467        let name = truncate_cstr::<{ libc::MAXTHREADNAMESIZE }>(name);
468        let res = libc::pthread_setname_np(name.as_ptr());
469        // We have no good way of propagating errors here, but in debug-builds let's check that this actually worked.
470        debug_assert_eq!(res, 0);
471    }
472}
473
474#[cfg(target_os = "netbsd")]
475pub fn set_name(name: &CStr) {
476    unsafe {
477        let res = libc::pthread_setname_np(
478            libc::pthread_self(),
479            c"%s".as_ptr(),
480            name.as_ptr() as *mut libc::c_void,
481        );
482        debug_assert_eq!(res, 0);
483    }
484}
485
486#[cfg(any(target_os = "solaris", target_os = "illumos", target_os = "nto"))]
487pub fn set_name(name: &CStr) {
488    weak!(
489        fn pthread_setname_np(thread: libc::pthread_t, name: *const libc::c_char) -> libc::c_int;
490    );
491
492    if let Some(f) = pthread_setname_np.get() {
493        #[cfg(target_os = "nto")]
494        const THREAD_NAME_MAX: usize = libc::_NTO_THREAD_NAME_MAX as usize;
495        #[cfg(any(target_os = "solaris", target_os = "illumos"))]
496        const THREAD_NAME_MAX: usize = 32;
497
498        let name = truncate_cstr::<{ THREAD_NAME_MAX }>(name);
499        let res = unsafe { f(libc::pthread_self(), name.as_ptr()) };
500        debug_assert_eq!(res, 0);
501    }
502}
503
504#[cfg(target_os = "fuchsia")]
505pub fn set_name(name: &CStr) {
506    use crate::sys::pal::fuchsia::*;
507    unsafe {
508        zx_object_set_property(
509            zx_thread_self(),
510            ZX_PROP_NAME,
511            name.as_ptr() as *const libc::c_void,
512            name.to_bytes().len(),
513        );
514    }
515}
516
517#[cfg(target_os = "haiku")]
518pub fn set_name(name: &CStr) {
519    unsafe {
520        let thread_self = libc::find_thread(ptr::null_mut());
521        let res = libc::rename_thread(thread_self, name.as_ptr());
522        // We have no good way of propagating errors here, but in debug-builds let's check that this actually worked.
523        debug_assert_eq!(res, libc::B_OK);
524    }
525}
526
527#[cfg(target_os = "vxworks")]
528pub fn set_name(name: &CStr) {
529    let mut name = truncate_cstr::<{ (libc::VX_TASK_RENAME_LENGTH - 1) as usize }>(name);
530    let res = unsafe { libc::taskNameSet(libc::taskIdSelf(), name.as_mut_ptr()) };
531    debug_assert_eq!(res, libc::OK);
532}
533
534#[cfg(not(any(target_os = "espidf", target_os = "wasi")))]
535pub fn sleep(dur: Duration) {
536    let mut secs = dur.as_secs();
537    let mut nsecs = dur.subsec_nanos() as _;
538
539    // If we're awoken with a signal then the return value will be -1 and
540    // nanosleep will fill in `ts` with the remaining time.
541    unsafe {
542        while secs > 0 || nsecs > 0 {
543            let mut ts = libc::timespec {
544                tv_sec: cmp::min(libc::time_t::MAX as u64, secs) as libc::time_t,
545                tv_nsec: nsecs,
546            };
547            secs -= ts.tv_sec as u64;
548            let ts_ptr = &raw mut ts;
549            if libc::nanosleep(ts_ptr, ts_ptr) == -1 {
550                assert_eq!(sys::io::errno(), libc::EINTR);
551                secs += ts.tv_sec as u64;
552                nsecs = ts.tv_nsec;
553            } else {
554                nsecs = 0;
555            }
556        }
557    }
558}
559
560#[cfg(any(
561    target_os = "espidf",
562    // wasi-libc prior to WebAssembly/wasi-libc#696 has a broken implementation
563    // of `nanosleep`, used above by most platforms, so use `usleep` until
564    // that fix propagates throughout the ecosystem.
565    target_os = "wasi",
566))]
567pub fn sleep(dur: Duration) {
568    // ESP-IDF does not have `nanosleep`, so we use `usleep` instead.
569    // As per the documentation of `usleep`, it is expected to support
570    // sleep times as big as at least up to 1 second.
571    //
572    // ESP-IDF does support almost up to `u32::MAX`, but due to a potential integer overflow in its
573    // `usleep` implementation
574    // (https://github.com/espressif/esp-idf/blob/d7ca8b94c852052e3bc33292287ef4dd62c9eeb1/components/newlib/time.c#L210),
575    // we limit the sleep time to the maximum one that would not cause the underlying `usleep` implementation to overflow
576    // (`portTICK_PERIOD_MS` can be anything between 1 to 1000, and is 10 by default).
577    const MAX_MICROS: u32 = u32::MAX - 1_000_000 - 1;
578
579    // Add any nanoseconds smaller than a microsecond as an extra microsecond
580    // so as to comply with the `std::thread::sleep` contract which mandates
581    // implementations to sleep for _at least_ the provided `dur`.
582    // We can't overflow `micros` as it is a `u128`, while `Duration` is a pair of
583    // (`u64` secs, `u32` nanos), where the nanos are strictly smaller than 1 second
584    // (i.e. < 1_000_000_000)
585    let mut micros = dur.as_micros() + if dur.subsec_nanos() % 1_000 > 0 { 1 } else { 0 };
586
587    while micros > 0 {
588        let st = if micros > MAX_MICROS as u128 { MAX_MICROS } else { micros as u32 };
589        unsafe {
590            libc::usleep(st);
591        }
592
593        micros -= st as u128;
594    }
595}
596
597// Any unix that has clock_nanosleep
598// If this list changes update the MIRI chock_nanosleep shim
599#[cfg(any(
600    target_os = "freebsd",
601    target_os = "netbsd",
602    target_os = "linux",
603    target_os = "android",
604    target_os = "solaris",
605    target_os = "illumos",
606    target_os = "dragonfly",
607    target_os = "hurd",
608    target_os = "fuchsia",
609    target_os = "vxworks",
610    target_os = "wasi",
611))]
612pub fn sleep_until(deadline: crate::time::Instant) {
613    use crate::time::Instant;
614
615    let Some(ts) = deadline.into_inner().into_timespec().to_timespec() else {
616        // The deadline is further in the future then can be passed to
617        // clock_nanosleep. We have to use Self::sleep instead. This might
618        // happen on 32 bit platforms, especially closer to 2038.
619        let now = Instant::now();
620        if let Some(delay) = deadline.checked_duration_since(now) {
621            sleep(delay);
622        }
623        return;
624    };
625
626    unsafe {
627        // When we get interrupted (res = EINTR) call clock_nanosleep again
628        loop {
629            let res = libc::clock_nanosleep(
630                crate::sys::time::Instant::CLOCK_ID,
631                libc::TIMER_ABSTIME,
632                &ts,
633                core::ptr::null_mut(), // not required with TIMER_ABSTIME
634            );
635
636            if res == 0 {
637                break;
638            } else {
639                assert_eq!(
640                    res,
641                    libc::EINTR,
642                    "timespec is in range,
643                         clockid is valid and kernel should support it"
644                );
645            }
646        }
647    }
648}
649
650pub fn yield_now() {
651    let ret = unsafe { libc::sched_yield() };
652    debug_assert_eq!(ret, 0);
653}
654
655#[cfg(any(target_os = "android", target_os = "linux"))]
656mod cgroups {
657    //! Currently not covered
658    //! * cgroup v2 in non-standard mountpoints
659    //! * paths containing control characters or spaces, since those would be escaped in procfs
660    //!   output and we don't unescape
661
662    use crate::borrow::Cow;
663    use crate::ffi::OsString;
664    use crate::fs::{File, exists};
665    use crate::io::{BufRead, Read};
666    use crate::os::unix::ffi::OsStringExt;
667    use crate::path::{Path, PathBuf};
668    use crate::str::from_utf8;
669
670    #[derive(PartialEq)]
671    enum Cgroup {
672        V1,
673        V2,
674    }
675
676    /// Returns cgroup CPU quota in core-equivalents, rounded down or usize::MAX if the quota cannot
677    /// be determined or is not set.
678    pub(super) fn quota() -> usize {
679        let mut quota = usize::MAX;
680        if cfg!(miri) {
681            // Attempting to open a file fails under default flags due to isolation.
682            // And Miri does not have parallelism anyway.
683            return quota;
684        }
685
686        let _: Option<()> = try {
687            let mut buf = Vec::with_capacity(128);
688            // find our place in the cgroup hierarchy
689            File::open("/proc/self/cgroup").ok()?.read_to_end(&mut buf).ok()?;
690            let (cgroup_path, version) =
691                buf.split(|&c| c == b'\n').fold(None, |previous, line| {
692                    let mut fields = line.splitn(3, |&c| c == b':');
693                    // 2nd field is a list of controllers for v1 or empty for v2
694                    let version = match fields.nth(1) {
695                        Some(b"") => Cgroup::V2,
696                        Some(controllers)
697                            if from_utf8(controllers)
698                                .is_ok_and(|c| c.split(',').any(|c| c == "cpu")) =>
699                        {
700                            Cgroup::V1
701                        }
702                        _ => return previous,
703                    };
704
705                    // already-found v1 trumps v2 since it explicitly specifies its controllers
706                    if previous.is_some() && version == Cgroup::V2 {
707                        return previous;
708                    }
709
710                    let path = fields.last()?;
711                    // skip leading slash
712                    Some((path[1..].to_owned(), version))
713                })?;
714            let cgroup_path = PathBuf::from(OsString::from_vec(cgroup_path));
715
716            quota = match version {
717                Cgroup::V1 => quota_v1(cgroup_path),
718                Cgroup::V2 => quota_v2(cgroup_path),
719            };
720        };
721
722        quota
723    }
724
725    fn quota_v2(group_path: PathBuf) -> usize {
726        let mut quota = usize::MAX;
727
728        let mut path = PathBuf::with_capacity(128);
729        let mut read_buf = String::with_capacity(20);
730
731        // standard mount location defined in file-hierarchy(7) manpage
732        let cgroup_mount = "/sys/fs/cgroup";
733
734        path.push(cgroup_mount);
735        path.push(&group_path);
736
737        path.push("cgroup.controllers");
738
739        // skip if we're not looking at cgroup2
740        if matches!(exists(&path), Err(_) | Ok(false)) {
741            return usize::MAX;
742        };
743
744        path.pop();
745
746        let _: Option<()> = try {
747            while path.starts_with(cgroup_mount) {
748                path.push("cpu.max");
749
750                read_buf.clear();
751
752                if File::open(&path).and_then(|mut f| f.read_to_string(&mut read_buf)).is_ok() {
753                    let raw_quota = read_buf.lines().next()?;
754                    let mut raw_quota = raw_quota.split(' ');
755                    let limit = raw_quota.next()?;
756                    let period = raw_quota.next()?;
757                    match (limit.parse::<usize>(), period.parse::<usize>()) {
758                        (Ok(limit), Ok(period)) if period > 0 => {
759                            quota = quota.min(limit / period);
760                        }
761                        _ => {}
762                    }
763                }
764
765                path.pop(); // pop filename
766                path.pop(); // pop dir
767            }
768        };
769
770        quota
771    }
772
773    fn quota_v1(group_path: PathBuf) -> usize {
774        let mut quota = usize::MAX;
775        let mut path = PathBuf::with_capacity(128);
776        let mut read_buf = String::with_capacity(20);
777
778        // Hardcode commonly used locations mentioned in the cgroups(7) manpage
779        // if that doesn't work scan mountinfo and adjust `group_path` for bind-mounts
780        let mounts: &[fn(&Path) -> Option<(_, &Path)>] = &[
781            |p| Some((Cow::Borrowed("/sys/fs/cgroup/cpu"), p)),
782            |p| Some((Cow::Borrowed("/sys/fs/cgroup/cpu,cpuacct"), p)),
783            // this can be expensive on systems with tons of mountpoints
784            // but we only get to this point when /proc/self/cgroups explicitly indicated
785            // this process belongs to a cpu-controller cgroup v1 and the defaults didn't work
786            find_mountpoint,
787        ];
788
789        for mount in mounts {
790            let Some((mount, group_path)) = mount(&group_path) else { continue };
791
792            path.clear();
793            path.push(mount.as_ref());
794            path.push(&group_path);
795
796            // skip if we guessed the mount incorrectly
797            if matches!(exists(&path), Err(_) | Ok(false)) {
798                continue;
799            }
800
801            while path.starts_with(mount.as_ref()) {
802                let mut parse_file = |name| {
803                    path.push(name);
804                    read_buf.clear();
805
806                    let f = File::open(&path);
807                    path.pop(); // restore buffer before any early returns
808                    f.ok()?.read_to_string(&mut read_buf).ok()?;
809                    let parsed = read_buf.trim().parse::<usize>().ok()?;
810
811                    Some(parsed)
812                };
813
814                let limit = parse_file("cpu.cfs_quota_us");
815                let period = parse_file("cpu.cfs_period_us");
816
817                match (limit, period) {
818                    (Some(limit), Some(period)) if period > 0 => quota = quota.min(limit / period),
819                    _ => {}
820                }
821
822                path.pop();
823            }
824
825            // we passed the try_exists above so we should have traversed the correct hierarchy
826            // when reaching this line
827            break;
828        }
829
830        quota
831    }
832
833    /// Scan mountinfo for cgroup v1 mountpoint with a cpu controller
834    ///
835    /// If the cgroupfs is a bind mount then `group_path` is adjusted to skip
836    /// over the already-included prefix
837    fn find_mountpoint(group_path: &Path) -> Option<(Cow<'static, str>, &Path)> {
838        let mut reader = File::open_buffered("/proc/self/mountinfo").ok()?;
839        let mut line = String::with_capacity(256);
840        loop {
841            line.clear();
842            if reader.read_line(&mut line).ok()? == 0 {
843                break;
844            }
845
846            let line = line.trim();
847            let mut items = line.split(' ');
848
849            let sub_path = items.nth(3)?;
850            let mount_point = items.next()?;
851            let mount_opts = items.next_back()?;
852            let filesystem_type = items.nth_back(1)?;
853
854            if filesystem_type != "cgroup" || !mount_opts.split(',').any(|opt| opt == "cpu") {
855                // not a cgroup / not a cpu-controller
856                continue;
857            }
858
859            let sub_path = Path::new(sub_path).strip_prefix("/").ok()?;
860
861            if !group_path.starts_with(sub_path) {
862                // this is a bind-mount and the bound subdirectory
863                // does not contain the cgroup this process belongs to
864                continue;
865            }
866
867            let trimmed_group_path = group_path.strip_prefix(sub_path).ok()?;
868
869            return Some((Cow::Owned(mount_point.to_owned()), trimmed_group_path));
870        }
871
872        None
873    }
874}
875
876// glibc >= 2.15 has a __pthread_get_minstack() function that returns
877// PTHREAD_STACK_MIN plus bytes needed for thread-local storage.
878// We need that information to avoid blowing up when a small stack
879// is created in an application with big thread-local storage requirements.
880// See #6233 for rationale and details.
881#[cfg(all(target_os = "linux", target_env = "gnu"))]
882unsafe fn min_stack_size(attr: *const libc::pthread_attr_t) -> usize {
883    // We use dlsym to avoid an ELF version dependency on GLIBC_PRIVATE. (#23628)
884    // We shouldn't really be using such an internal symbol, but there's currently
885    // no other way to account for the TLS size.
886    dlsym!(
887        fn __pthread_get_minstack(attr: *const libc::pthread_attr_t) -> libc::size_t;
888    );
889
890    match __pthread_get_minstack.get() {
891        None => libc::PTHREAD_STACK_MIN,
892        Some(f) => unsafe { f(attr) },
893    }
894}
895
896// No point in looking up __pthread_get_minstack() on non-glibc platforms.
897#[cfg(all(
898    not(all(target_os = "linux", target_env = "gnu")),
899    not(any(target_os = "netbsd", target_os = "nuttx"))
900))]
901unsafe fn min_stack_size(_: *const libc::pthread_attr_t) -> usize {
902    libc::PTHREAD_STACK_MIN
903}
904
905#[cfg(any(target_os = "netbsd", target_os = "nuttx"))]
906unsafe fn min_stack_size(_: *const libc::pthread_attr_t) -> usize {
907    static STACK: crate::sync::OnceLock<usize> = crate::sync::OnceLock::new();
908
909    *STACK.get_or_init(|| {
910        let mut stack = unsafe { libc::sysconf(libc::_SC_THREAD_STACK_MIN) };
911        if stack < 0 {
912            stack = 2048; // just a guess
913        }
914
915        stack as usize
916    })
917}
std/sys/thread/unix.rs

std/sys/thread/
unix.rs
