std/sys/thread/

unix.rs

#[cfg(not(any(
    target_env = "newlib",
    target_os = "l4re",
    target_os = "emscripten",
    target_os = "redox",
    target_os = "hurd",
    target_os = "aix",
    target_os = "wasi",
)))]
use crate::ffi::CStr;
use crate::mem::{self, DropGuard, ManuallyDrop};
use crate::num::NonZero;
#[cfg(all(target_os = "linux", target_env = "gnu"))]
use crate::sys::weak::dlsym;
#[cfg(any(target_os = "solaris", target_os = "illumos", target_os = "nto",))]
use crate::sys::weak::weak;
use crate::thread::ThreadInit;
use crate::time::Duration;
use crate::{cmp, io, ptr, sys};
#[cfg(not(any(
    target_os = "l4re",
    target_os = "vxworks",
    target_os = "espidf",
    target_os = "nuttx"
)))]
pub const DEFAULT_MIN_STACK_SIZE: usize = 2 * 1024 * 1024;
#[cfg(target_os = "l4re")]
pub const DEFAULT_MIN_STACK_SIZE: usize = 1024 * 1024;
#[cfg(target_os = "vxworks")]
pub const DEFAULT_MIN_STACK_SIZE: usize = 256 * 1024;
#[cfg(any(target_os = "espidf", target_os = "nuttx"))]
pub const DEFAULT_MIN_STACK_SIZE: usize = 0; // 0 indicates that the stack size configured in the ESP-IDF/NuttX menuconfig system should be used

pub struct Thread {
    id: libc::pthread_t,
}

// Some platforms may have pthread_t as a pointer in which case we still want
// a thread to be Send/Sync
unsafe impl Send for Thread {}
unsafe impl Sync for Thread {}

impl Thread {
    // unsafe: see thread::Builder::spawn_unchecked for safety requirements
    #[cfg_attr(miri, track_caller)] // even without panics, this helps for Miri backtraces
    pub unsafe fn new(stack: usize, init: Box<ThreadInit>) -> io::Result<Thread> {
        // FIXME: remove this block once wasi-sdk is updated with the fix from
        // https://github.com/WebAssembly/wasi-libc/pull/716
        // WASI does not support threading via pthreads. While wasi-libc provides
        // pthread stubs, pthread_create returns EAGAIN, which causes confusing
        // errors. We return UNSUPPORTED_PLATFORM directly instead.
        if cfg!(all(target_os = "wasi", not(target_feature = "atomics"))) {
            return Err(io::Error::UNSUPPORTED_PLATFORM);
        }

        let data = init;
        let mut attr: mem::MaybeUninit<libc::pthread_attr_t> = mem::MaybeUninit::uninit();
        assert_eq!(libc::pthread_attr_init(attr.as_mut_ptr()), 0);
        let mut attr = DropGuard::new(&mut attr, |attr| {
            assert_eq!(libc::pthread_attr_destroy(attr.as_mut_ptr()), 0)
        });

        #[cfg(any(target_os = "espidf", target_os = "nuttx"))]
        if stack > 0 {
            // Only set the stack if a non-zero value is passed
            // 0 is used as an indication that the default stack size configured in the ESP-IDF/NuttX menuconfig system should be used
            assert_eq!(
                libc::pthread_attr_setstacksize(
                    attr.as_mut_ptr(),
                    cmp::max(stack, min_stack_size(attr.as_ptr()))
                ),
                0
            );
        }

        #[cfg(not(any(target_os = "espidf", target_os = "nuttx")))]
        {
            let stack_size = cmp::max(stack, min_stack_size(attr.as_ptr()));

            match libc::pthread_attr_setstacksize(attr.as_mut_ptr(), stack_size) {
                0 => {}
                n => {
                    assert_eq!(n, libc::EINVAL);
                    // EINVAL means |stack_size| is either too small or not a
                    // multiple of the system page size. Because it's definitely
                    // >= PTHREAD_STACK_MIN, it must be an alignment issue.
                    // Round up to the nearest page and try again.
                    let page_size = sys::os::page_size();
                    let stack_size =
                        (stack_size + page_size - 1) & (-(page_size as isize - 1) as usize - 1);

                    // Some libc implementations, e.g. musl, place an upper bound
                    // on the stack size, in which case we can only gracefully return
                    // an error here.
                    if libc::pthread_attr_setstacksize(attr.as_mut_ptr(), stack_size) != 0 {
                        return Err(io::const_error!(
                            io::ErrorKind::InvalidInput,
                            "invalid stack size"
                        ));
                    }
                }
            };
        }

        let data = Box::into_raw(data);
        let mut native: libc::pthread_t = mem::zeroed();
        let ret = libc::pthread_create(&mut native, attr.as_ptr(), thread_start, data as *mut _);
        return if ret == 0 {
            Ok(Thread { id: native })
        } else {
            // The thread failed to start and as a result `data` was not consumed.
            // Therefore, it is safe to reconstruct the box so that it gets deallocated.
            drop(Box::from_raw(data));
            Err(io::Error::from_raw_os_error(ret))
        };

        extern "C" fn thread_start(data: *mut libc::c_void) -> *mut libc::c_void {
            unsafe {
                // SAFETY: we are simply recreating the box that was leaked earlier.
                let init = Box::from_raw(data as *mut ThreadInit);
                let rust_start = init.init();

                // Now that the thread information is set, set up our stack
                // overflow handler.
                let _handler = sys::stack_overflow::Handler::new();

                rust_start();
            }
            ptr::null_mut()
        }
    }

    pub fn join(self) {
        let id = self.into_id();
        let ret = unsafe { libc::pthread_join(id, ptr::null_mut()) };
        assert!(ret == 0, "failed to join thread: {}", io::Error::from_raw_os_error(ret));
    }

    #[cfg(not(target_os = "wasi"))]
    pub fn id(&self) -> libc::pthread_t {
        self.id
    }

    pub fn into_id(self) -> libc::pthread_t {
        ManuallyDrop::new(self).id
    }
}

impl Drop for Thread {
    fn drop(&mut self) {
        let ret = unsafe { libc::pthread_detach(self.id) };
        debug_assert_eq!(ret, 0);
    }
}

pub fn available_parallelism() -> io::Result<NonZero<usize>> {
    cfg_select! {
        any(
            target_os = "android",
            target_os = "emscripten",
            target_os = "fuchsia",
            target_os = "hurd",
            target_os = "linux",
            target_os = "aix",
            target_vendor = "apple",
            target_os = "cygwin",
        ) => {
            #[allow(unused_assignments)]
            #[allow(unused_mut)]
            let mut quota = usize::MAX;

            #[cfg(any(target_os = "android", target_os = "linux"))]
            {
                quota = cgroups::quota().max(1);
                let mut set: libc::cpu_set_t = unsafe { mem::zeroed() };
                unsafe {
                    if libc::sched_getaffinity(0, size_of::<libc::cpu_set_t>(), &mut set) == 0 {
                        let count = libc::CPU_COUNT(&set) as usize;
                        let count = count.min(quota);

                        // According to sched_getaffinity's API it should always be non-zero, but
                        // some old MIPS kernels were buggy and zero-initialized the mask if
                        // none was explicitly set.
                        // In that case we use the sysconf fallback.
                        if let Some(count) = NonZero::new(count) {
                            return Ok(count)
                        }
                    }
                }
            }
            match unsafe { libc::sysconf(libc::_SC_NPROCESSORS_ONLN) } {
                -1 => Err(io::Error::last_os_error()),
                0 => Err(io::Error::UNKNOWN_THREAD_COUNT),
                cpus => {
                    let count = cpus as usize;
                    // Cover the unusual situation where we were able to get the quota but not the affinity mask
                    let count = count.min(quota);
                    Ok(unsafe { NonZero::new_unchecked(count) })
                }
            }
        }
        any(
           target_os = "freebsd",
           target_os = "dragonfly",
           target_os = "openbsd",
           target_os = "netbsd",
        ) => {
            use crate::ptr;

            #[cfg(target_os = "freebsd")]
            {
                let mut set: libc::cpuset_t = unsafe { mem::zeroed() };
                unsafe {
                    if libc::cpuset_getaffinity(
                        libc::CPU_LEVEL_WHICH,
                        libc::CPU_WHICH_PID,
                        -1,
                        size_of::<libc::cpuset_t>(),
                        &mut set,
                    ) == 0 {
                        let count = libc::CPU_COUNT(&set) as usize;
                        if count > 0 {
                            return Ok(NonZero::new_unchecked(count));
                        }
                    }
                }
            }

            #[cfg(target_os = "netbsd")]
            {
                unsafe {
                    let set = libc::_cpuset_create();
                    if !set.is_null() {
                        let mut count: usize = 0;
                        if libc::pthread_getaffinity_np(libc::pthread_self(), libc::_cpuset_size(set), set) == 0 {
                            for i in 0..libc::cpuid_t::MAX {
                                match libc::_cpuset_isset(i, set) {
                                    -1 => break,
                                    0 => continue,
                                    _ => count = count + 1,
                                }
                            }
                        }
                        libc::_cpuset_destroy(set);
                        if let Some(count) = NonZero::new(count) {
                            return Ok(count);
                        }
                    }
                }
            }

            let mut cpus: libc::c_uint = 0;
            let mut cpus_size = size_of_val(&cpus);

            unsafe {
                cpus = libc::sysconf(libc::_SC_NPROCESSORS_ONLN) as libc::c_uint;
            }

            // Fallback approach in case of errors or no hardware threads.
            if cpus < 1 {
                let mut mib = [libc::CTL_HW, libc::HW_NCPU, 0, 0];
                let res = unsafe {
                    libc::sysctl(
                        mib.as_mut_ptr(),
                        2,
                        (&raw mut cpus) as *mut _,
                        (&raw mut cpus_size) as *mut _,
                        ptr::null_mut(),
                        0,
                    )
                };

                // Handle errors if any.
                if res == -1 {
                    return Err(io::Error::last_os_error());
                } else if cpus == 0 {
                    return Err(io::Error::UNKNOWN_THREAD_COUNT);
                }
            }

            Ok(unsafe { NonZero::new_unchecked(cpus as usize) })
        }
        target_os = "nto" => {
            unsafe {
                use libc::_syspage_ptr;
                if _syspage_ptr.is_null() {
                    Err(io::const_error!(io::ErrorKind::NotFound, "no syspage available"))
                } else {
                    let cpus = (*_syspage_ptr).num_cpu;
                    NonZero::new(cpus as usize)
                        .ok_or(io::Error::UNKNOWN_THREAD_COUNT)
                }
            }
        }
        any(target_os = "solaris", target_os = "illumos") => {
            let mut cpus = 0u32;
            if unsafe { libc::pset_info(libc::PS_MYID, core::ptr::null_mut(), &mut cpus, core::ptr::null_mut()) } != 0 {
                return Err(io::Error::UNKNOWN_THREAD_COUNT);
            }
            Ok(unsafe { NonZero::new_unchecked(cpus as usize) })
        }
        target_os = "haiku" => {
            // system_info cpu_count field gets the static data set at boot time with `smp_set_num_cpus`
            // `get_system_info` calls then `smp_get_num_cpus`
            unsafe {
                let mut sinfo: libc::system_info = crate::mem::zeroed();
                let res = libc::get_system_info(&mut sinfo);

                if res != libc::B_OK {
                    return Err(io::Error::UNKNOWN_THREAD_COUNT);
                }

                Ok(NonZero::new_unchecked(sinfo.cpu_count as usize))
            }
        }
        target_os = "vxworks" => {
            // Note: there is also `vxCpuConfiguredGet`, closer to _SC_NPROCESSORS_CONF
            // expectations than the actual cores availability.

            // SAFETY: `vxCpuEnabledGet` always fetches a mask with at least one bit set
            unsafe{
                let set = libc::vxCpuEnabledGet();
                Ok(NonZero::new_unchecked(set.count_ones() as usize))
            }
        }
        _ => {
            // FIXME: implement on Redox, l4re
            Err(io::const_error!(io::ErrorKind::Unsupported, "getting the number of hardware threads is not supported on the target platform"))
        }
    }
}

pub fn current_os_id() -> Option<u64> {
    // Most Unix platforms have a way to query an integer ID of the current thread, all with
    // slightly different spellings.
    //
    // The OS thread ID is used rather than `pthread_self` so as to match what will be displayed
    // for process inspection (debuggers, trace, `top`, etc.).
    cfg_select! {
        // Most platforms have a function returning a `pid_t` or int, which is an `i32`.
        any(target_os = "android", target_os = "linux") => {
            use crate::sys::pal::weak::syscall;

            // `libc::gettid` is only available on glibc 2.30+, but the syscall is available
            // since Linux 2.4.11.
            syscall!(fn gettid() -> libc::pid_t;);

            // SAFETY: FFI call with no preconditions.
            let id: libc::pid_t = unsafe { gettid() };
            Some(id as u64)
        }
        target_os = "nto" => {
            // SAFETY: FFI call with no preconditions.
            let id: libc::pid_t = unsafe { libc::gettid() };
            Some(id as u64)
        }
        target_os = "openbsd" => {
            // SAFETY: FFI call with no preconditions.
            let id: libc::pid_t = unsafe { libc::getthrid() };
            Some(id as u64)
        }
        target_os = "freebsd" => {
            // SAFETY: FFI call with no preconditions.
            let id: libc::c_int = unsafe { libc::pthread_getthreadid_np() };
            Some(id as u64)
        }
        target_os = "netbsd" => {
            // SAFETY: FFI call with no preconditions.
            let id: libc::lwpid_t = unsafe { libc::_lwp_self() };
            Some(id as u64)
        }
        any(target_os = "illumos", target_os = "solaris") => {
            // On Illumos and Solaris, the `pthread_t` is the same as the OS thread ID.
            // SAFETY: FFI call with no preconditions.
            let id: libc::pthread_t = unsafe { libc::pthread_self() };
            Some(id as u64)
        }
        target_vendor = "apple" => {
            // Apple allows querying arbitrary thread IDs, `thread=NULL` queries the current thread.
            let mut id = 0u64;
            // SAFETY: `thread_id` is a valid pointer, no other preconditions.
            let status: libc::c_int = unsafe { libc::pthread_threadid_np(0, &mut id) };
            if status == 0 {
                Some(id)
            } else {
                None
            }
        }
        // Other platforms don't have an OS thread ID or don't have a way to access it.
        _ => None,
    }
}

#[cfg(any(
    target_os = "linux",
    target_os = "nto",
    target_os = "solaris",
    target_os = "illumos",
    target_os = "vxworks",
    target_os = "cygwin",
    target_vendor = "apple",
))]
fn truncate_cstr<const MAX_WITH_NUL: usize>(cstr: &CStr) -> [libc::c_char; MAX_WITH_NUL] {
    let mut result = [0; MAX_WITH_NUL];
    for (src, dst) in cstr.to_bytes().iter().zip(&mut result[..MAX_WITH_NUL - 1]) {
        *dst = *src as libc::c_char;
    }
    result
}

#[cfg(target_os = "android")]
pub fn set_name(name: &CStr) {
    const PR_SET_NAME: libc::c_int = 15;
    unsafe {
        let res = libc::prctl(
            PR_SET_NAME,
            name.as_ptr(),
            0 as libc::c_ulong,
            0 as libc::c_ulong,
            0 as libc::c_ulong,
        );
        // We have no good way of propagating errors here, but in debug-builds let's check that this actually worked.
        debug_assert_eq!(res, 0);
    }
}

#[cfg(any(
    target_os = "linux",
    target_os = "freebsd",
    target_os = "dragonfly",
    target_os = "nuttx",
    target_os = "cygwin"
))]
pub fn set_name(name: &CStr) {
    unsafe {
        cfg_select! {
            any(target_os = "linux", target_os = "cygwin") => {
                // Linux and Cygwin limits the allowed length of the name.
                const TASK_COMM_LEN: usize = 16;
                let name = truncate_cstr::<{ TASK_COMM_LEN }>(name);
            }
            _ => {
                // FreeBSD, DragonFly BSD and NuttX do not enforce length limits.
            }
        };
        // Available since glibc 2.12, musl 1.1.16, and uClibc 1.0.20 for Linux,
        // FreeBSD 12.2 and 13.0, and DragonFly BSD 6.0.
        let res = libc::pthread_setname_np(libc::pthread_self(), name.as_ptr());
        // We have no good way of propagating errors here, but in debug-builds let's check that this actually worked.
        debug_assert_eq!(res, 0);
    }
}

#[cfg(target_os = "openbsd")]
pub fn set_name(name: &CStr) {
    unsafe {
        libc::pthread_set_name_np(libc::pthread_self(), name.as_ptr());
    }
}

#[cfg(target_vendor = "apple")]
pub fn set_name(name: &CStr) {
    unsafe {
        let name = truncate_cstr::<{ libc::MAXTHREADNAMESIZE }>(name);
        let res = libc::pthread_setname_np(name.as_ptr());
        // We have no good way of propagating errors here, but in debug-builds let's check that this actually worked.
        debug_assert_eq!(res, 0);
    }
}

#[cfg(target_os = "netbsd")]
pub fn set_name(name: &CStr) {
    unsafe {
        let res = libc::pthread_setname_np(
            libc::pthread_self(),
            c"%s".as_ptr(),
            name.as_ptr() as *mut libc::c_void,
        );
        debug_assert_eq!(res, 0);
    }
}

#[cfg(any(target_os = "solaris", target_os = "illumos", target_os = "nto"))]
pub fn set_name(name: &CStr) {
    weak!(
        fn pthread_setname_np(thread: libc::pthread_t, name: *const libc::c_char) -> libc::c_int;
    );

    if let Some(f) = pthread_setname_np.get() {
        #[cfg(target_os = "nto")]
        const THREAD_NAME_MAX: usize = libc::_NTO_THREAD_NAME_MAX as usize;
        #[cfg(any(target_os = "solaris", target_os = "illumos"))]
        const THREAD_NAME_MAX: usize = 32;

        let name = truncate_cstr::<{ THREAD_NAME_MAX }>(name);
        let res = unsafe { f(libc::pthread_self(), name.as_ptr()) };
        debug_assert_eq!(res, 0);
    }
}

#[cfg(target_os = "fuchsia")]
pub fn set_name(name: &CStr) {
    use crate::sys::pal::fuchsia::*;
    unsafe {
        zx_object_set_property(
            zx_thread_self(),
            ZX_PROP_NAME,
            name.as_ptr() as *const libc::c_void,
            name.to_bytes().len(),
        );
    }
}

#[cfg(target_os = "haiku")]
pub fn set_name(name: &CStr) {
    unsafe {
        let thread_self = libc::find_thread(ptr::null_mut());
        let res = libc::rename_thread(thread_self, name.as_ptr());
        // We have no good way of propagating errors here, but in debug-builds let's check that this actually worked.
        debug_assert_eq!(res, libc::B_OK);
    }
}

#[cfg(target_os = "vxworks")]
pub fn set_name(name: &CStr) {
    let mut name = truncate_cstr::<{ (libc::VX_TASK_RENAME_LENGTH - 1) as usize }>(name);
    let res = unsafe { libc::taskNameSet(libc::taskIdSelf(), name.as_mut_ptr()) };
    debug_assert_eq!(res, libc::OK);
}

#[cfg(not(target_os = "espidf"))]
pub fn sleep(dur: Duration) {
    cfg_select! {
        // Any unix that has clock_nanosleep
        // If this list changes update the MIRI chock_nanosleep shim
        any(
            target_os = "freebsd",
            target_os = "netbsd",
            target_os = "linux",
            target_os = "android",
            target_os = "solaris",
            target_os = "illumos",
            target_os = "dragonfly",
            target_os = "hurd",
            target_os = "vxworks",
            target_os = "wasi",
        ) => {
            // POSIX specifies that `nanosleep` uses CLOCK_REALTIME, but is not
            // affected by clock adjustments. The timing of `sleep` however should
            // be tied to `Instant` where possible. Thus, we use `clock_nanosleep`
            // with a relative time interval instead, which allows explicitly
            // specifying the clock.
            //
            // In practice, most systems (like e.g. Linux) actually use
            // CLOCK_MONOTONIC for `nanosleep` anyway, but others like FreeBSD don't
            // so it's better to be safe.
            //
            // wasi-libc prior to WebAssembly/wasi-libc#696 has a broken implementation
            // of `nanosleep` which used `CLOCK_REALTIME` even though it is unsupported
            // on WASIp2. Using `clock_nanosleep` directly bypasses the issue.
            unsafe fn nanosleep(rqtp: *const libc::timespec, rmtp: *mut libc::timespec) -> libc::c_int {
                unsafe { libc::clock_nanosleep(crate::sys::time::Instant::CLOCK_ID, 0, rqtp, rmtp) }
            }
        }
        _ => {
            unsafe fn nanosleep(rqtp: *const libc::timespec, rmtp: *mut libc::timespec) -> libc::c_int {
                let r = unsafe { libc::nanosleep(rqtp, rmtp) };
                // `clock_nanosleep` returns the error number directly, so mimic
                // that behaviour to make the shared code below simpler.
                if r == 0 { 0 } else { sys::io::errno() }
            }
        }
    }

    let mut secs = dur.as_secs();
    let mut nsecs = dur.subsec_nanos() as _;

    // If we're awoken with a signal then the return value will be -1 and
    // nanosleep will fill in `ts` with the remaining time.
    unsafe {
        while secs > 0 || nsecs > 0 {
            let mut ts = libc::timespec {
                tv_sec: cmp::min(libc::time_t::MAX as u64, secs) as libc::time_t,
                tv_nsec: nsecs,
            };
            secs -= ts.tv_sec as u64;
            let ts_ptr = &raw mut ts;
            let r = nanosleep(ts_ptr, ts_ptr);
            if r != 0 {
                assert_eq!(r, libc::EINTR);
                secs += ts.tv_sec as u64;
                nsecs = ts.tv_nsec;
            } else {
                nsecs = 0;
            }
        }
    }
}

#[cfg(target_os = "espidf")]
pub fn sleep(dur: Duration) {
    // ESP-IDF does not have `nanosleep`, so we use `usleep` instead.
    // As per the documentation of `usleep`, it is expected to support
    // sleep times as big as at least up to 1 second.
    //
    // ESP-IDF does support almost up to `u32::MAX`, but due to a potential integer overflow in its
    // `usleep` implementation
    // (https://github.com/espressif/esp-idf/blob/d7ca8b94c852052e3bc33292287ef4dd62c9eeb1/components/newlib/time.c#L210),
    // we limit the sleep time to the maximum one that would not cause the underlying `usleep` implementation to overflow
    // (`portTICK_PERIOD_MS` can be anything between 1 to 1000, and is 10 by default).
    const MAX_MICROS: u32 = u32::MAX - 1_000_000 - 1;

    // Add any nanoseconds smaller than a microsecond as an extra microsecond
    // so as to comply with the `std::thread::sleep` contract which mandates
    // implementations to sleep for _at least_ the provided `dur`.
    // We can't overflow `micros` as it is a `u128`, while `Duration` is a pair of
    // (`u64` secs, `u32` nanos), where the nanos are strictly smaller than 1 second
    // (i.e. < 1_000_000_000)
    let mut micros = dur.as_micros() + if dur.subsec_nanos() % 1_000 > 0 { 1 } else { 0 };

    while micros > 0 {
        let st = if micros > MAX_MICROS as u128 { MAX_MICROS } else { micros as u32 };
        unsafe {
            libc::usleep(st);
        }

        micros -= st as u128;
    }
}

// Any unix that has clock_nanosleep
// If this list changes update the MIRI chock_nanosleep shim
#[cfg(any(
    target_os = "freebsd",
    target_os = "netbsd",
    target_os = "linux",
    target_os = "android",
    target_os = "solaris",
    target_os = "illumos",
    target_os = "dragonfly",
    target_os = "hurd",
    target_os = "vxworks",
    target_os = "wasi",
))]
pub fn sleep_until(deadline: crate::time::Instant) {
    use crate::time::Instant;

    #[cfg(all(
        target_os = "linux",
        target_env = "gnu",
        target_pointer_width = "32",
        not(target_arch = "riscv32")
    ))]
    {
        use crate::sys::pal::time::__timespec64;
        use crate::sys::pal::weak::weak;

        // This got added in glibc 2.31, along with a 64-bit `clock_gettime`
        // function.
        weak! {
            fn __clock_nanosleep_time64(
                clock_id: libc::clockid_t,
                flags: libc::c_int,
                req: *const __timespec64,
                rem: *mut __timespec64,
            ) -> libc::c_int;
        }

        if let Some(clock_nanosleep) = __clock_nanosleep_time64.get() {
            let ts = deadline.into_inner().into_timespec().to_timespec64();
            loop {
                let r = unsafe {
                    clock_nanosleep(
                        crate::sys::time::Instant::CLOCK_ID,
                        libc::TIMER_ABSTIME,
                        &ts,
                        core::ptr::null_mut(),
                    )
                };

                match r {
                    0 => return,
                    libc::EINTR => continue,
                    // If the underlying kernel doesn't support the 64-bit
                    // syscall, `__clock_nanosleep_time64` will fail. The
                    // error code nowadays is EOVERFLOW, but it used to be
                    // ENOSYS – so just don't rely on any particular value.
                    // The parameters are all valid, so the only reasons
                    // why the call might fail are EINTR and the call not
                    // being supported. Fall through to the clamping version
                    // in that case.
                    _ => break,
                }
            }
        }
    }

    let Some(ts) = deadline.into_inner().into_timespec().to_timespec() else {
        // The deadline is further in the future then can be passed to
        // clock_nanosleep. We have to use Self::sleep instead. This might
        // happen on 32 bit platforms, especially closer to 2038.
        let now = Instant::now();
        if let Some(delay) = deadline.checked_duration_since(now) {
            sleep(delay);
        }
        return;
    };

    unsafe {
        // When we get interrupted (res = EINTR) call clock_nanosleep again
        loop {
            let res = libc::clock_nanosleep(
                crate::sys::time::Instant::CLOCK_ID,
                libc::TIMER_ABSTIME,
                &ts,
                core::ptr::null_mut(), // not required with TIMER_ABSTIME
            );

            if res == 0 {
                break;
            } else {
                assert_eq!(
                    res,
                    libc::EINTR,
                    "timespec is in range,
                         clockid is valid and kernel should support it"
                );
            }
        }
    }
}

#[cfg(target_vendor = "apple")]
pub fn sleep_until(deadline: crate::time::Instant) {
    unsafe extern "C" {
        // This is defined in the public header mach/mach_time.h alongside
        // `mach_absolute_time`, and like it has been available since the very
        // beginning.
        //
        // There isn't really any documentation on this function, except for a
        // short reference in technical note 2169:
        // https://developer.apple.com/library/archive/technotes/tn2169/_index.html
        safe fn mach_wait_until(deadline: u64) -> libc::kern_return_t;
    }

    // Make sure to round up to ensure that we definitely sleep until after
    // the deadline has elapsed.
    let Some(deadline) = deadline.into_inner().into_mach_absolute_time_ceil() else {
        // Since the deadline is before the system boot time, it has already
        // passed, so we can return immediately.
        return;
    };

    // If the deadline is not representable, then sleep for the maximum duration
    // possible and worry about the potential clock issues later (in ca. 600 years).
    let deadline = deadline.try_into().unwrap_or(u64::MAX);
    loop {
        match mach_wait_until(deadline) {
            // Success! The deadline has passed.
            libc::KERN_SUCCESS => break,
            // If the sleep gets interrupted by a signal, `mach_wait_until`
            // returns KERN_ABORTED, so we need to restart the syscall.
            // Also see Apple's implementation of the POSIX `nanosleep`, which
            // converts this error to the POSIX equivalent EINTR:
            // https://github.com/apple-oss-distributions/Libc/blob/55b54c0a0c37b3b24393b42b90a4c561d6c606b1/gen/nanosleep.c#L281-L306
            libc::KERN_ABORTED => continue,
            // All other errors indicate that something has gone wrong...
            error => {
                let description = unsafe { CStr::from_ptr(libc::mach_error_string(error)) };
                panic!("mach_wait_until failed: {} (code {error})", description.display())
            }
        }
    }
}

pub fn yield_now() {
    let ret = unsafe { libc::sched_yield() };
    debug_assert_eq!(ret, 0);
}

#[cfg(any(target_os = "android", target_os = "linux"))]
mod cgroups {
    //! Currently not covered
    //! * cgroup v2 in non-standard mountpoints
    //! * paths containing control characters or spaces, since those would be escaped in procfs
    //!   output and we don't unescape

    use crate::borrow::Cow;
    use crate::ffi::OsString;
    use crate::fs::{File, exists};
    use crate::io::{BufRead, Read};
    use crate::os::unix::ffi::OsStringExt;
    use crate::path::{Path, PathBuf};
    use crate::str::from_utf8;

    #[derive(PartialEq)]
    enum Cgroup {
        V1,
        V2,
    }

    /// Returns cgroup CPU quota in core-equivalents, rounded down or usize::MAX if the quota cannot
    /// be determined or is not set.
    pub(super) fn quota() -> usize {
        let mut quota = usize::MAX;
        if cfg!(miri) {
            // Attempting to open a file fails under default flags due to isolation.
            // And Miri does not have parallelism anyway.
            return quota;
        }

        let _: Option<()> = try {
            let mut buf = Vec::with_capacity(128);
            // find our place in the cgroup hierarchy
            File::open("/proc/self/cgroup").ok()?.read_to_end(&mut buf).ok()?;
            let (cgroup_path, version) =
                buf.split(|&c| c == b'\n').fold(None, |previous, line| {
                    let mut fields = line.splitn(3, |&c| c == b':');
                    // 2nd field is a list of controllers for v1 or empty for v2
                    let version = match fields.nth(1) {
                        Some(b"") => Cgroup::V2,
                        Some(controllers)
                            if from_utf8(controllers)
                                .is_ok_and(|c| c.split(',').any(|c| c == "cpu")) =>
                        {
                            Cgroup::V1
                        }
                        _ => return previous,
                    };

                    // already-found v1 trumps v2 since it explicitly specifies its controllers
                    if previous.is_some() && version == Cgroup::V2 {
                        return previous;
                    }

                    let path = fields.last()?;
                    // skip leading slash
                    Some((path[1..].to_owned(), version))
                })?;
            let cgroup_path = PathBuf::from(OsString::from_vec(cgroup_path));

            quota = match version {
                Cgroup::V1 => quota_v1(cgroup_path),
                Cgroup::V2 => quota_v2(cgroup_path),
            };
        };

        quota
    }

    fn quota_v2(group_path: PathBuf) -> usize {
        let mut quota = usize::MAX;

        let mut path = PathBuf::with_capacity(128);
        let mut read_buf = String::with_capacity(20);

        // standard mount location defined in file-hierarchy(7) manpage
        let cgroup_mount = "/sys/fs/cgroup";

        path.push(cgroup_mount);
        path.push(&group_path);

        path.push("cgroup.controllers");

        // skip if we're not looking at cgroup2
        if matches!(exists(&path), Err(_) | Ok(false)) {
            return usize::MAX;
        };

        path.pop();

        let _: Option<()> = try {
            while path.starts_with(cgroup_mount) {
                path.push("cpu.max");

                read_buf.clear();

                if File::open(&path).and_then(|mut f| f.read_to_string(&mut read_buf)).is_ok() {
                    let raw_quota = read_buf.lines().next()?;
                    let mut raw_quota = raw_quota.split(' ');
                    let limit = raw_quota.next()?;
                    let period = raw_quota.next()?;
                    match (limit.parse::<usize>(), period.parse::<usize>()) {
                        (Ok(limit), Ok(period)) if period > 0 => {
                            quota = quota.min(limit / period);
                        }
                        _ => {}
                    }
                }

                path.pop(); // pop filename
                path.pop(); // pop dir
            }
        };

        quota
    }

    fn quota_v1(group_path: PathBuf) -> usize {
        let mut quota = usize::MAX;
        let mut path = PathBuf::with_capacity(128);
        let mut read_buf = String::with_capacity(20);

        // Hardcode commonly used locations mentioned in the cgroups(7) manpage
        // if that doesn't work scan mountinfo and adjust `group_path` for bind-mounts
        let mounts: &[fn(&Path) -> Option<(_, &Path)>] = &[
            |p| Some((Cow::Borrowed("/sys/fs/cgroup/cpu"), p)),
            |p| Some((Cow::Borrowed("/sys/fs/cgroup/cpu,cpuacct"), p)),
            // this can be expensive on systems with tons of mountpoints
            // but we only get to this point when /proc/self/cgroups explicitly indicated
            // this process belongs to a cpu-controller cgroup v1 and the defaults didn't work
            find_mountpoint,
        ];

        for mount in mounts {
            let Some((mount, group_path)) = mount(&group_path) else { continue };

            path.clear();
            path.push(mount.as_ref());
            path.push(&group_path);

            // skip if we guessed the mount incorrectly
            if matches!(exists(&path), Err(_) | Ok(false)) {
                continue;
            }

            while path.starts_with(mount.as_ref()) {
                let mut parse_file = |name| {
                    path.push(name);
                    read_buf.clear();

                    let f = File::open(&path);
                    path.pop(); // restore buffer before any early returns
                    f.ok()?.read_to_string(&mut read_buf).ok()?;
                    let parsed = read_buf.trim().parse::<usize>().ok()?;

                    Some(parsed)
                };

                let limit = parse_file("cpu.cfs_quota_us");
                let period = parse_file("cpu.cfs_period_us");

                match (limit, period) {
                    (Some(limit), Some(period)) if period > 0 => quota = quota.min(limit / period),
                    _ => {}
                }

                path.pop();
            }

            // we passed the try_exists above so we should have traversed the correct hierarchy
            // when reaching this line
            break;
        }

        quota
    }

    /// Scan mountinfo for cgroup v1 mountpoint with a cpu controller
    ///
    /// If the cgroupfs is a bind mount then `group_path` is adjusted to skip
    /// over the already-included prefix
    fn find_mountpoint(group_path: &Path) -> Option<(Cow<'static, str>, &Path)> {
        let mut reader = File::open_buffered("/proc/self/mountinfo").ok()?;
        let mut line = String::with_capacity(256);
        loop {
            line.clear();
            if reader.read_line(&mut line).ok()? == 0 {
                break;
            }

            let line = line.trim();
            let mut items = line.split(' ');

            let sub_path = items.nth(3)?;
            let mount_point = items.next()?;
            let mount_opts = items.next_back()?;
            let filesystem_type = items.nth_back(1)?;

            if filesystem_type != "cgroup" || !mount_opts.split(',').any(|opt| opt == "cpu") {
                // not a cgroup / not a cpu-controller
                continue;
            }

            let sub_path = Path::new(sub_path).strip_prefix("/").ok()?;

            if !group_path.starts_with(sub_path) {
                // this is a bind-mount and the bound subdirectory
                // does not contain the cgroup this process belongs to
                continue;
            }

            let trimmed_group_path = group_path.strip_prefix(sub_path).ok()?;

            return Some((Cow::Owned(mount_point.to_owned()), trimmed_group_path));
        }

        None
    }
}

// glibc >= 2.15 has a __pthread_get_minstack() function that returns
// PTHREAD_STACK_MIN plus bytes needed for thread-local storage.
// We need that information to avoid blowing up when a small stack
// is created in an application with big thread-local storage requirements.
// See #6233 for rationale and details.
#[cfg(all(target_os = "linux", target_env = "gnu"))]
unsafe fn min_stack_size(attr: *const libc::pthread_attr_t) -> usize {
    // We use dlsym to avoid an ELF version dependency on GLIBC_PRIVATE. (#23628)
    // We shouldn't really be using such an internal symbol, but there's currently
    // no other way to account for the TLS size.
    dlsym!(
        fn __pthread_get_minstack(attr: *const libc::pthread_attr_t) -> libc::size_t;
    );

    match __pthread_get_minstack.get() {
        None => libc::PTHREAD_STACK_MIN,
        Some(f) => unsafe { f(attr) },
    }
}

// No point in looking up __pthread_get_minstack() on non-glibc platforms.
#[cfg(all(
    not(all(target_os = "linux", target_env = "gnu")),
    not(any(target_os = "netbsd", target_os = "nuttx"))
))]
unsafe fn min_stack_size(_: *const libc::pthread_attr_t) -> usize {
    libc::PTHREAD_STACK_MIN
}

#[cfg(any(target_os = "netbsd", target_os = "nuttx"))]
unsafe fn min_stack_size(_: *const libc::pthread_attr_t) -> usize {
    static STACK: crate::sync::OnceLock<usize> = crate::sync::OnceLock::new();

    *STACK.get_or_init(|| {
        let mut stack = unsafe { libc::sysconf(libc::_SC_THREAD_STACK_MIN) };
        if stack < 0 {
            stack = 2048; // just a guess
        }

        stack as usize
    })
}