1 | use crate::ffi::CStr; |
2 | use crate::mem::{self, ManuallyDrop}; |
3 | use crate::num::NonZero; |
4 | #[cfg (all(target_os = "linux" , target_env = "gnu" ))] |
5 | use crate::sys::weak::dlsym; |
6 | #[cfg (any(target_os = "solaris" , target_os = "illumos" , target_os = "nto" ,))] |
7 | use crate::sys::weak::weak; |
8 | use crate::sys::{os, stack_overflow}; |
9 | use crate::time::Duration; |
10 | use crate::{cmp, io, ptr}; |
11 | #[cfg (not(any(target_os = "l4re" , target_os = "vxworks" , target_os = "espidf" )))] |
12 | pub const DEFAULT_MIN_STACK_SIZE: usize = 2 * 1024 * 1024; |
13 | #[cfg (target_os = "l4re" )] |
14 | pub const DEFAULT_MIN_STACK_SIZE: usize = 1024 * 1024; |
15 | #[cfg (target_os = "vxworks" )] |
16 | pub const DEFAULT_MIN_STACK_SIZE: usize = 256 * 1024; |
17 | #[cfg (target_os = "espidf" )] |
18 | pub 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" )] |
21 | mod 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 | |
37 | pub 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 |
43 | unsafe impl Send for Thread {} |
44 | unsafe impl Sync for Thread {} |
45 | |
46 | impl 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 | target_os = "cygwin" |
142 | ))] |
143 | pub fn set_name(name: &CStr) { |
144 | unsafe { |
145 | cfg_if::cfg_if! { |
146 | if #[cfg(any(target_os = "linux" , target_os = "cygwin" ))] { |
147 | // Linux and Cygwin limits the allowed length of the name. |
148 | const TASK_COMM_LEN: usize = 16; |
149 | let name = truncate_cstr::<{ TASK_COMM_LEN }>(name); |
150 | } else { |
151 | // FreeBSD, DragonFly BSD and NuttX do not enforce length limits. |
152 | } |
153 | }; |
154 | // Available since glibc 2.12, musl 1.1.16, and uClibc 1.0.20 for Linux, |
155 | // FreeBSD 12.2 and 13.0, and DragonFly BSD 6.0. |
156 | let res = libc::pthread_setname_np(libc::pthread_self(), name.as_ptr()); |
157 | // We have no good way of propagating errors here, but in debug-builds let's check that this actually worked. |
158 | debug_assert_eq!(res, 0); |
159 | } |
160 | } |
161 | |
162 | #[cfg (target_os = "openbsd" )] |
163 | pub fn set_name(name: &CStr) { |
164 | unsafe { |
165 | libc::pthread_set_name_np(libc::pthread_self(), name.as_ptr()); |
166 | } |
167 | } |
168 | |
169 | #[cfg (target_vendor = "apple" )] |
170 | pub fn set_name(name: &CStr) { |
171 | unsafe { |
172 | let name = truncate_cstr::<{ libc::MAXTHREADNAMESIZE }>(name); |
173 | let res = libc::pthread_setname_np(name.as_ptr()); |
174 | // We have no good way of propagating errors here, but in debug-builds let's check that this actually worked. |
175 | debug_assert_eq!(res, 0); |
176 | } |
177 | } |
178 | |
179 | #[cfg (target_os = "netbsd" )] |
180 | pub fn set_name(name: &CStr) { |
181 | unsafe { |
182 | let res = libc::pthread_setname_np( |
183 | libc::pthread_self(), |
184 | c"%s" .as_ptr(), |
185 | name.as_ptr() as *mut libc::c_void, |
186 | ); |
187 | debug_assert_eq!(res, 0); |
188 | } |
189 | } |
190 | |
191 | #[cfg (any(target_os = "solaris" , target_os = "illumos" , target_os = "nto" ))] |
192 | // FIXME(#115199): Rust currently omits weak function definitions |
193 | // and its metadata from LLVM IR. |
194 | #[no_sanitize (cfi)] |
195 | pub fn set_name(name: &CStr) { |
196 | weak! { |
197 | fn pthread_setname_np( |
198 | libc::pthread_t, *const libc::c_char |
199 | ) -> libc::c_int |
200 | } |
201 | |
202 | if let Some(f) = pthread_setname_np.get() { |
203 | #[cfg (target_os = "nto" )] |
204 | const THREAD_NAME_MAX: usize = libc::_NTO_THREAD_NAME_MAX as usize; |
205 | #[cfg (any(target_os = "solaris" , target_os = "illumos" ))] |
206 | const THREAD_NAME_MAX: usize = 32; |
207 | |
208 | let name = truncate_cstr::<{ THREAD_NAME_MAX }>(name); |
209 | let res = unsafe { f(libc::pthread_self(), name.as_ptr()) }; |
210 | debug_assert_eq!(res, 0); |
211 | } |
212 | } |
213 | |
214 | #[cfg (target_os = "fuchsia" )] |
215 | pub fn set_name(name: &CStr) { |
216 | use self::zircon::*; |
217 | unsafe { |
218 | zx_object_set_property( |
219 | zx_thread_self(), |
220 | ZX_PROP_NAME, |
221 | name.as_ptr() as *const libc::c_void, |
222 | name.to_bytes().len(), |
223 | ); |
224 | } |
225 | } |
226 | |
227 | #[cfg (target_os = "haiku" )] |
228 | pub fn set_name(name: &CStr) { |
229 | unsafe { |
230 | let thread_self = libc::find_thread(ptr::null_mut()); |
231 | let res = libc::rename_thread(thread_self, name.as_ptr()); |
232 | // We have no good way of propagating errors here, but in debug-builds let's check that this actually worked. |
233 | debug_assert_eq!(res, libc::B_OK); |
234 | } |
235 | } |
236 | |
237 | #[cfg (target_os = "vxworks" )] |
238 | pub fn set_name(name: &CStr) { |
239 | // FIXME(libc): adding real STATUS, ERROR type eventually. |
240 | unsafe extern "C" { |
241 | fn taskNameSet(task_id: libc::TASK_ID, task_name: *mut libc::c_char) -> libc::c_int; |
242 | } |
243 | |
244 | // VX_TASK_NAME_LEN is 31 in VxWorks 7. |
245 | const VX_TASK_NAME_LEN: usize = 31; |
246 | |
247 | let mut name = truncate_cstr::<{ VX_TASK_NAME_LEN }>(name); |
248 | let res = unsafe { taskNameSet(libc::taskIdSelf(), name.as_mut_ptr()) }; |
249 | debug_assert_eq!(res, libc::OK); |
250 | } |
251 | |
252 | #[cfg (any( |
253 | target_env = "newlib" , |
254 | target_os = "l4re" , |
255 | target_os = "emscripten" , |
256 | target_os = "redox" , |
257 | target_os = "hurd" , |
258 | target_os = "aix" , |
259 | ))] |
260 | pub fn set_name(_name: &CStr) { |
261 | // Newlib and Emscripten have no way to set a thread name. |
262 | } |
263 | |
264 | #[cfg (not(target_os = "espidf" ))] |
265 | pub fn sleep(dur: Duration) { |
266 | let mut secs = dur.as_secs(); |
267 | let mut nsecs = dur.subsec_nanos() as _; |
268 | |
269 | // If we're awoken with a signal then the return value will be -1 and |
270 | // nanosleep will fill in `ts` with the remaining time. |
271 | unsafe { |
272 | while secs > 0 || nsecs > 0 { |
273 | let mut ts = libc::timespec { |
274 | tv_sec: cmp::min(libc::time_t::MAX as u64, secs) as libc::time_t, |
275 | tv_nsec: nsecs, |
276 | }; |
277 | secs -= ts.tv_sec as u64; |
278 | let ts_ptr = &raw mut ts; |
279 | if libc::nanosleep(ts_ptr, ts_ptr) == -1 { |
280 | assert_eq!(os::errno(), libc::EINTR); |
281 | secs += ts.tv_sec as u64; |
282 | nsecs = ts.tv_nsec; |
283 | } else { |
284 | nsecs = 0; |
285 | } |
286 | } |
287 | } |
288 | } |
289 | |
290 | #[cfg (target_os = "espidf" )] |
291 | pub fn sleep(dur: Duration) { |
292 | // ESP-IDF does not have `nanosleep`, so we use `usleep` instead. |
293 | // As per the documentation of `usleep`, it is expected to support |
294 | // sleep times as big as at least up to 1 second. |
295 | // |
296 | // ESP-IDF does support almost up to `u32::MAX`, but due to a potential integer overflow in its |
297 | // `usleep` implementation |
298 | // (https://github.com/espressif/esp-idf/blob/d7ca8b94c852052e3bc33292287ef4dd62c9eeb1/components/newlib/time.c#L210), |
299 | // we limit the sleep time to the maximum one that would not cause the underlying `usleep` implementation to overflow |
300 | // (`portTICK_PERIOD_MS` can be anything between 1 to 1000, and is 10 by default). |
301 | const MAX_MICROS: u32 = u32::MAX - 1_000_000 - 1; |
302 | |
303 | // Add any nanoseconds smaller than a microsecond as an extra microsecond |
304 | // so as to comply with the `std::thread::sleep` contract which mandates |
305 | // implementations to sleep for _at least_ the provided `dur`. |
306 | // We can't overflow `micros` as it is a `u128`, while `Duration` is a pair of |
307 | // (`u64` secs, `u32` nanos), where the nanos are strictly smaller than 1 second |
308 | // (i.e. < 1_000_000_000) |
309 | let mut micros = dur.as_micros() + if dur.subsec_nanos() % 1_000 > 0 { 1 } else { 0 }; |
310 | |
311 | while micros > 0 { |
312 | let st = if micros > MAX_MICROS as u128 { MAX_MICROS } else { micros as u32 }; |
313 | unsafe { |
314 | libc::usleep(st); |
315 | } |
316 | |
317 | micros -= st as u128; |
318 | } |
319 | } |
320 | |
321 | pub fn join(self) { |
322 | let id = self.into_id(); |
323 | let ret = unsafe { libc::pthread_join(id, ptr::null_mut()) }; |
324 | assert!(ret == 0, "failed to join thread: {}" , io::Error::from_raw_os_error(ret)); |
325 | } |
326 | |
327 | pub fn id(&self) -> libc::pthread_t { |
328 | self.id |
329 | } |
330 | |
331 | pub fn into_id(self) -> libc::pthread_t { |
332 | ManuallyDrop::new(self).id |
333 | } |
334 | } |
335 | |
336 | impl Drop for Thread { |
337 | fn drop(&mut self) { |
338 | let ret: i32 = unsafe { libc::pthread_detach(self.id) }; |
339 | debug_assert_eq!(ret, 0); |
340 | } |
341 | } |
342 | |
343 | #[cfg (any( |
344 | target_os = "linux" , |
345 | target_os = "nto" , |
346 | target_os = "solaris" , |
347 | target_os = "illumos" , |
348 | target_os = "vxworks" , |
349 | target_os = "cygwin" , |
350 | target_vendor = "apple" , |
351 | ))] |
352 | fn truncate_cstr<const MAX_WITH_NUL: usize>(cstr: &CStr) -> [libc::c_char; MAX_WITH_NUL] { |
353 | let mut result: [i8; MAX_WITH_NUL] = [0; MAX_WITH_NUL]; |
354 | for (src: &u8, dst: &mut i8) in cstr.to_bytes().iter().zip(&mut result[..MAX_WITH_NUL - 1]) { |
355 | *dst = *src as libc::c_char; |
356 | } |
357 | result |
358 | } |
359 | |
360 | pub fn available_parallelism() -> io::Result<NonZero<usize>> { |
361 | cfg_if::cfg_if! { |
362 | if #[cfg(any( |
363 | target_os = "android" , |
364 | target_os = "emscripten" , |
365 | target_os = "fuchsia" , |
366 | target_os = "hurd" , |
367 | target_os = "linux" , |
368 | target_os = "aix" , |
369 | target_vendor = "apple" , |
370 | target_os = "cygwin" , |
371 | ))] { |
372 | #[allow (unused_assignments)] |
373 | #[allow (unused_mut)] |
374 | let mut quota = usize::MAX; |
375 | |
376 | #[cfg (any(target_os = "android" , target_os = "linux" ))] |
377 | { |
378 | quota = cgroups::quota().max(1); |
379 | let mut set: libc::cpu_set_t = unsafe { mem::zeroed() }; |
380 | unsafe { |
381 | if libc::sched_getaffinity(0, size_of::<libc::cpu_set_t>(), &mut set) == 0 { |
382 | let count = libc::CPU_COUNT(&set) as usize; |
383 | let count = count.min(quota); |
384 | |
385 | // According to sched_getaffinity's API it should always be non-zero, but |
386 | // some old MIPS kernels were buggy and zero-initialized the mask if |
387 | // none was explicitly set. |
388 | // In that case we use the sysconf fallback. |
389 | if let Some(count) = NonZero::new(count) { |
390 | return Ok(count) |
391 | } |
392 | } |
393 | } |
394 | } |
395 | match unsafe { libc::sysconf(libc::_SC_NPROCESSORS_ONLN) } { |
396 | -1 => Err(io::Error::last_os_error()), |
397 | 0 => Err(io::Error::UNKNOWN_THREAD_COUNT), |
398 | cpus => { |
399 | let count = cpus as usize; |
400 | // Cover the unusual situation where we were able to get the quota but not the affinity mask |
401 | let count = count.min(quota); |
402 | Ok(unsafe { NonZero::new_unchecked(count) }) |
403 | } |
404 | } |
405 | } else if #[cfg(any( |
406 | target_os = "freebsd" , |
407 | target_os = "dragonfly" , |
408 | target_os = "openbsd" , |
409 | target_os = "netbsd" , |
410 | ))] { |
411 | use crate::ptr; |
412 | |
413 | #[cfg (target_os = "freebsd" )] |
414 | { |
415 | let mut set: libc::cpuset_t = unsafe { mem::zeroed() }; |
416 | unsafe { |
417 | if libc::cpuset_getaffinity( |
418 | libc::CPU_LEVEL_WHICH, |
419 | libc::CPU_WHICH_PID, |
420 | -1, |
421 | size_of::<libc::cpuset_t>(), |
422 | &mut set, |
423 | ) == 0 { |
424 | let count = libc::CPU_COUNT(&set) as usize; |
425 | if count > 0 { |
426 | return Ok(NonZero::new_unchecked(count)); |
427 | } |
428 | } |
429 | } |
430 | } |
431 | |
432 | #[cfg (target_os = "netbsd" )] |
433 | { |
434 | unsafe { |
435 | let set = libc::_cpuset_create(); |
436 | if !set.is_null() { |
437 | let mut count: usize = 0; |
438 | if libc::pthread_getaffinity_np(libc::pthread_self(), libc::_cpuset_size(set), set) == 0 { |
439 | for i in 0..libc::cpuid_t::MAX { |
440 | match libc::_cpuset_isset(i, set) { |
441 | -1 => break, |
442 | 0 => continue, |
443 | _ => count = count + 1, |
444 | } |
445 | } |
446 | } |
447 | libc::_cpuset_destroy(set); |
448 | if let Some(count) = NonZero::new(count) { |
449 | return Ok(count); |
450 | } |
451 | } |
452 | } |
453 | } |
454 | |
455 | let mut cpus: libc::c_uint = 0; |
456 | let mut cpus_size = size_of_val(&cpus); |
457 | |
458 | unsafe { |
459 | cpus = libc::sysconf(libc::_SC_NPROCESSORS_ONLN) as libc::c_uint; |
460 | } |
461 | |
462 | // Fallback approach in case of errors or no hardware threads. |
463 | if cpus < 1 { |
464 | let mut mib = [libc::CTL_HW, libc::HW_NCPU, 0, 0]; |
465 | let res = unsafe { |
466 | libc::sysctl( |
467 | mib.as_mut_ptr(), |
468 | 2, |
469 | (&raw mut cpus) as *mut _, |
470 | (&raw mut cpus_size) as *mut _, |
471 | ptr::null_mut(), |
472 | 0, |
473 | ) |
474 | }; |
475 | |
476 | // Handle errors if any. |
477 | if res == -1 { |
478 | return Err(io::Error::last_os_error()); |
479 | } else if cpus == 0 { |
480 | return Err(io::Error::UNKNOWN_THREAD_COUNT); |
481 | } |
482 | } |
483 | |
484 | Ok(unsafe { NonZero::new_unchecked(cpus as usize) }) |
485 | } else if #[cfg(target_os = "nto" )] { |
486 | unsafe { |
487 | use libc::_syspage_ptr; |
488 | if _syspage_ptr.is_null() { |
489 | Err(io::const_error!(io::ErrorKind::NotFound, "no syspage available" )) |
490 | } else { |
491 | let cpus = (*_syspage_ptr).num_cpu; |
492 | NonZero::new(cpus as usize) |
493 | .ok_or(io::Error::UNKNOWN_THREAD_COUNT) |
494 | } |
495 | } |
496 | } else if #[cfg(any(target_os = "solaris" , target_os = "illumos" ))] { |
497 | let mut cpus = 0u32; |
498 | if unsafe { libc::pset_info(libc::PS_MYID, core::ptr::null_mut(), &mut cpus, core::ptr::null_mut()) } != 0 { |
499 | return Err(io::Error::UNKNOWN_THREAD_COUNT); |
500 | } |
501 | Ok(unsafe { NonZero::new_unchecked(cpus as usize) }) |
502 | } else if #[cfg(target_os = "haiku" )] { |
503 | // system_info cpu_count field gets the static data set at boot time with `smp_set_num_cpus` |
504 | // `get_system_info` calls then `smp_get_num_cpus` |
505 | unsafe { |
506 | let mut sinfo: libc::system_info = crate::mem::zeroed(); |
507 | let res = libc::get_system_info(&mut sinfo); |
508 | |
509 | if res != libc::B_OK { |
510 | return Err(io::Error::UNKNOWN_THREAD_COUNT); |
511 | } |
512 | |
513 | Ok(NonZero::new_unchecked(sinfo.cpu_count as usize)) |
514 | } |
515 | } else if #[cfg(target_os = "vxworks" )] { |
516 | // Note: there is also `vxCpuConfiguredGet`, closer to _SC_NPROCESSORS_CONF |
517 | // expectations than the actual cores availability. |
518 | unsafe extern "C" { |
519 | fn vxCpuEnabledGet() -> libc::cpuset_t; |
520 | } |
521 | |
522 | // SAFETY: `vxCpuEnabledGet` always fetches a mask with at least one bit set |
523 | unsafe{ |
524 | let set = vxCpuEnabledGet(); |
525 | Ok(NonZero::new_unchecked(set.count_ones() as usize)) |
526 | } |
527 | } else { |
528 | // FIXME: implement on Redox, l4re |
529 | Err(io::const_error!(io::ErrorKind::Unsupported, "getting the number of hardware threads is not supported on the target platform" )) |
530 | } |
531 | } |
532 | } |
533 | |
534 | #[cfg (any(target_os = "android" , target_os = "linux" ))] |
535 | mod cgroups { |
536 | //! Currently not covered |
537 | //! * cgroup v2 in non-standard mountpoints |
538 | //! * paths containing control characters or spaces, since those would be escaped in procfs |
539 | //! output and we don't unescape |
540 | |
541 | use crate::borrow::Cow; |
542 | use crate::ffi::OsString; |
543 | use crate::fs::{File, exists}; |
544 | use crate::io::{BufRead, Read}; |
545 | use crate::os::unix::ffi::OsStringExt; |
546 | use crate::path::{Path, PathBuf}; |
547 | use crate::str::from_utf8; |
548 | |
549 | #[derive (PartialEq)] |
550 | enum Cgroup { |
551 | V1, |
552 | V2, |
553 | } |
554 | |
555 | /// Returns cgroup CPU quota in core-equivalents, rounded down or usize::MAX if the quota cannot |
556 | /// be determined or is not set. |
557 | pub(super) fn quota() -> usize { |
558 | let mut quota = usize::MAX; |
559 | if cfg!(miri) { |
560 | // Attempting to open a file fails under default flags due to isolation. |
561 | // And Miri does not have parallelism anyway. |
562 | return quota; |
563 | } |
564 | |
565 | let _: Option<()> = try { |
566 | let mut buf = Vec::with_capacity(128); |
567 | // find our place in the cgroup hierarchy |
568 | File::open("/proc/self/cgroup" ).ok()?.read_to_end(&mut buf).ok()?; |
569 | let (cgroup_path, version) = |
570 | buf.split(|&c| c == b' \n' ).fold(None, |previous, line| { |
571 | let mut fields = line.splitn(3, |&c| c == b':' ); |
572 | // 2nd field is a list of controllers for v1 or empty for v2 |
573 | let version = match fields.nth(1) { |
574 | Some(b"" ) => Cgroup::V2, |
575 | Some(controllers) |
576 | if from_utf8(controllers) |
577 | .is_ok_and(|c| c.split(',' ).any(|c| c == "cpu" )) => |
578 | { |
579 | Cgroup::V1 |
580 | } |
581 | _ => return previous, |
582 | }; |
583 | |
584 | // already-found v1 trumps v2 since it explicitly specifies its controllers |
585 | if previous.is_some() && version == Cgroup::V2 { |
586 | return previous; |
587 | } |
588 | |
589 | let path = fields.last()?; |
590 | // skip leading slash |
591 | Some((path[1..].to_owned(), version)) |
592 | })?; |
593 | let cgroup_path = PathBuf::from(OsString::from_vec(cgroup_path)); |
594 | |
595 | quota = match version { |
596 | Cgroup::V1 => quota_v1(cgroup_path), |
597 | Cgroup::V2 => quota_v2(cgroup_path), |
598 | }; |
599 | }; |
600 | |
601 | quota |
602 | } |
603 | |
604 | fn quota_v2(group_path: PathBuf) -> usize { |
605 | let mut quota = usize::MAX; |
606 | |
607 | let mut path = PathBuf::with_capacity(128); |
608 | let mut read_buf = String::with_capacity(20); |
609 | |
610 | // standard mount location defined in file-hierarchy(7) manpage |
611 | let cgroup_mount = "/sys/fs/cgroup" ; |
612 | |
613 | path.push(cgroup_mount); |
614 | path.push(&group_path); |
615 | |
616 | path.push("cgroup.controllers" ); |
617 | |
618 | // skip if we're not looking at cgroup2 |
619 | if matches!(exists(&path), Err(_) | Ok(false)) { |
620 | return usize::MAX; |
621 | }; |
622 | |
623 | path.pop(); |
624 | |
625 | let _: Option<()> = try { |
626 | while path.starts_with(cgroup_mount) { |
627 | path.push("cpu.max" ); |
628 | |
629 | read_buf.clear(); |
630 | |
631 | if File::open(&path).and_then(|mut f| f.read_to_string(&mut read_buf)).is_ok() { |
632 | let raw_quota = read_buf.lines().next()?; |
633 | let mut raw_quota = raw_quota.split(' ' ); |
634 | let limit = raw_quota.next()?; |
635 | let period = raw_quota.next()?; |
636 | match (limit.parse::<usize>(), period.parse::<usize>()) { |
637 | (Ok(limit), Ok(period)) if period > 0 => { |
638 | quota = quota.min(limit / period); |
639 | } |
640 | _ => {} |
641 | } |
642 | } |
643 | |
644 | path.pop(); // pop filename |
645 | path.pop(); // pop dir |
646 | } |
647 | }; |
648 | |
649 | quota |
650 | } |
651 | |
652 | fn quota_v1(group_path: PathBuf) -> usize { |
653 | let mut quota = usize::MAX; |
654 | let mut path = PathBuf::with_capacity(128); |
655 | let mut read_buf = String::with_capacity(20); |
656 | |
657 | // Hardcode commonly used locations mentioned in the cgroups(7) manpage |
658 | // if that doesn't work scan mountinfo and adjust `group_path` for bind-mounts |
659 | let mounts: &[fn(&Path) -> Option<(_, &Path)>] = &[ |
660 | |p| Some((Cow::Borrowed("/sys/fs/cgroup/cpu" ), p)), |
661 | |p| Some((Cow::Borrowed("/sys/fs/cgroup/cpu,cpuacct" ), p)), |
662 | // this can be expensive on systems with tons of mountpoints |
663 | // but we only get to this point when /proc/self/cgroups explicitly indicated |
664 | // this process belongs to a cpu-controller cgroup v1 and the defaults didn't work |
665 | find_mountpoint, |
666 | ]; |
667 | |
668 | for mount in mounts { |
669 | let Some((mount, group_path)) = mount(&group_path) else { continue }; |
670 | |
671 | path.clear(); |
672 | path.push(mount.as_ref()); |
673 | path.push(&group_path); |
674 | |
675 | // skip if we guessed the mount incorrectly |
676 | if matches!(exists(&path), Err(_) | Ok(false)) { |
677 | continue; |
678 | } |
679 | |
680 | while path.starts_with(mount.as_ref()) { |
681 | let mut parse_file = |name| { |
682 | path.push(name); |
683 | read_buf.clear(); |
684 | |
685 | let f = File::open(&path); |
686 | path.pop(); // restore buffer before any early returns |
687 | f.ok()?.read_to_string(&mut read_buf).ok()?; |
688 | let parsed = read_buf.trim().parse::<usize>().ok()?; |
689 | |
690 | Some(parsed) |
691 | }; |
692 | |
693 | let limit = parse_file("cpu.cfs_quota_us" ); |
694 | let period = parse_file("cpu.cfs_period_us" ); |
695 | |
696 | match (limit, period) { |
697 | (Some(limit), Some(period)) if period > 0 => quota = quota.min(limit / period), |
698 | _ => {} |
699 | } |
700 | |
701 | path.pop(); |
702 | } |
703 | |
704 | // we passed the try_exists above so we should have traversed the correct hierarchy |
705 | // when reaching this line |
706 | break; |
707 | } |
708 | |
709 | quota |
710 | } |
711 | |
712 | /// Scan mountinfo for cgroup v1 mountpoint with a cpu controller |
713 | /// |
714 | /// If the cgroupfs is a bind mount then `group_path` is adjusted to skip |
715 | /// over the already-included prefix |
716 | fn find_mountpoint(group_path: &Path) -> Option<(Cow<'static, str>, &Path)> { |
717 | let mut reader = File::open_buffered("/proc/self/mountinfo" ).ok()?; |
718 | let mut line = String::with_capacity(256); |
719 | loop { |
720 | line.clear(); |
721 | if reader.read_line(&mut line).ok()? == 0 { |
722 | break; |
723 | } |
724 | |
725 | let line = line.trim(); |
726 | let mut items = line.split(' ' ); |
727 | |
728 | let sub_path = items.nth(3)?; |
729 | let mount_point = items.next()?; |
730 | let mount_opts = items.next_back()?; |
731 | let filesystem_type = items.nth_back(1)?; |
732 | |
733 | if filesystem_type != "cgroup" || !mount_opts.split(',' ).any(|opt| opt == "cpu" ) { |
734 | // not a cgroup / not a cpu-controller |
735 | continue; |
736 | } |
737 | |
738 | let sub_path = Path::new(sub_path).strip_prefix("/" ).ok()?; |
739 | |
740 | if !group_path.starts_with(sub_path) { |
741 | // this is a bind-mount and the bound subdirectory |
742 | // does not contain the cgroup this process belongs to |
743 | continue; |
744 | } |
745 | |
746 | let trimmed_group_path = group_path.strip_prefix(sub_path).ok()?; |
747 | |
748 | return Some((Cow::Owned(mount_point.to_owned()), trimmed_group_path)); |
749 | } |
750 | |
751 | None |
752 | } |
753 | } |
754 | |
755 | // glibc >= 2.15 has a __pthread_get_minstack() function that returns |
756 | // PTHREAD_STACK_MIN plus bytes needed for thread-local storage. |
757 | // We need that information to avoid blowing up when a small stack |
758 | // is created in an application with big thread-local storage requirements. |
759 | // See #6233 for rationale and details. |
760 | #[cfg (all(target_os = "linux" , target_env = "gnu" ))] |
761 | unsafe fn min_stack_size(attr: *const libc::pthread_attr_t) -> usize { |
762 | // We use dlsym to avoid an ELF version dependency on GLIBC_PRIVATE. (#23628) |
763 | // We shouldn't really be using such an internal symbol, but there's currently |
764 | // no other way to account for the TLS size. |
765 | dlsym!(fn __pthread_get_minstack(*const libc::pthread_attr_t) -> libc::size_t); |
766 | |
767 | match __pthread_get_minstack.get() { |
768 | None => libc::PTHREAD_STACK_MIN, |
769 | Some(f) => unsafe { f(attr) }, |
770 | } |
771 | } |
772 | |
773 | // No point in looking up __pthread_get_minstack() on non-glibc platforms. |
774 | #[cfg (all( |
775 | not(all(target_os = "linux" , target_env = "gnu" )), |
776 | not(any(target_os = "netbsd" , target_os = "nuttx" )) |
777 | ))] |
778 | unsafe fn min_stack_size(_: *const libc::pthread_attr_t) -> usize { |
779 | libc::PTHREAD_STACK_MIN |
780 | } |
781 | |
782 | #[cfg (any(target_os = "netbsd" , target_os = "nuttx" ))] |
783 | unsafe fn min_stack_size(_: *const libc::pthread_attr_t) -> usize { |
784 | static STACK: crate::sync::OnceLock<usize> = crate::sync::OnceLock::new(); |
785 | |
786 | *STACK.get_or_init(|| { |
787 | let mut stack = unsafe { libc::sysconf(libc::_SC_THREAD_STACK_MIN) }; |
788 | if stack < 0 { |
789 | stack = 2048; // just a guess |
790 | } |
791 | |
792 | stack as usize |
793 | }) |
794 | } |
795 | |