1 | //===- Unix/Threading.inc - Unix Threading Implementation ----- -*- C++ -*-===// |
2 | // |
3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
4 | // See https://llvm.org/LICENSE.txt for license information. |
5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
6 | // |
7 | //===----------------------------------------------------------------------===// |
8 | // |
9 | // This file provides the Unix specific implementation of Threading functions. |
10 | // |
11 | //===----------------------------------------------------------------------===// |
12 | |
13 | #include "Unix.h" |
14 | #include "llvm/ADT/ScopeExit.h" |
15 | #include "llvm/ADT/SmallString.h" |
16 | #include "llvm/ADT/SmallVector.h" |
17 | #include "llvm/ADT/StringRef.h" |
18 | #include "llvm/ADT/Twine.h" |
19 | #include "llvm/Support/MemoryBuffer.h" |
20 | #include "llvm/Support/raw_ostream.h" |
21 | |
22 | #if defined(__APPLE__) |
23 | #include <mach/mach_init.h> |
24 | #include <mach/mach_port.h> |
25 | #include <pthread/qos.h> |
26 | #include <sys/sysctl.h> |
27 | #include <sys/types.h> |
28 | #endif |
29 | |
30 | #include <pthread.h> |
31 | |
32 | #if defined(__FreeBSD__) || defined(__OpenBSD__) |
33 | #include <pthread_np.h> // For pthread_getthreadid_np() / pthread_set_name_np() |
34 | #endif |
35 | |
36 | #if defined(__FreeBSD__) || defined(__FreeBSD_kernel__) |
37 | #include <errno.h> |
38 | #include <sys/cpuset.h> |
39 | #include <sys/sysctl.h> |
40 | #include <sys/user.h> |
41 | #include <unistd.h> |
42 | #endif |
43 | |
44 | #if defined(__NetBSD__) |
45 | #include <lwp.h> // For _lwp_self() |
46 | #endif |
47 | |
48 | #if defined(__OpenBSD__) |
49 | #include <unistd.h> // For getthrid() |
50 | #endif |
51 | |
52 | #if defined(__linux__) |
53 | #include <sched.h> // For sched_getaffinity |
54 | #include <sys/syscall.h> // For syscall codes |
55 | #include <unistd.h> // For syscall() |
56 | #endif |
57 | |
58 | namespace llvm { |
59 | pthread_t |
60 | llvm_execute_on_thread_impl(void *(*ThreadFunc)(void *), void *Arg, |
61 | std::optional<unsigned> StackSizeInBytes) { |
62 | int errnum; |
63 | |
64 | // Construct the attributes object. |
65 | pthread_attr_t Attr; |
66 | if ((errnum = ::pthread_attr_init(attr: &Attr)) != 0) { |
67 | ReportErrnumFatal(Msg: "pthread_attr_init failed" , errnum); |
68 | } |
69 | |
70 | auto AttrGuard = llvm::make_scope_exit(F: [&] { |
71 | if ((errnum = ::pthread_attr_destroy(attr: &Attr)) != 0) { |
72 | ReportErrnumFatal(Msg: "pthread_attr_destroy failed" , errnum); |
73 | } |
74 | }); |
75 | |
76 | // Set the requested stack size, if given. |
77 | if (StackSizeInBytes) { |
78 | if ((errnum = ::pthread_attr_setstacksize(attr: &Attr, stacksize: *StackSizeInBytes)) != 0) { |
79 | ReportErrnumFatal(Msg: "pthread_attr_setstacksize failed" , errnum); |
80 | } |
81 | } |
82 | |
83 | // Construct and execute the thread. |
84 | pthread_t Thread; |
85 | if ((errnum = ::pthread_create(newthread: &Thread, attr: &Attr, start_routine: ThreadFunc, arg: Arg)) != 0) |
86 | ReportErrnumFatal(Msg: "pthread_create failed" , errnum); |
87 | |
88 | return Thread; |
89 | } |
90 | |
91 | void llvm_thread_detach_impl(pthread_t Thread) { |
92 | int errnum; |
93 | |
94 | if ((errnum = ::pthread_detach(th: Thread)) != 0) { |
95 | ReportErrnumFatal(Msg: "pthread_detach failed" , errnum); |
96 | } |
97 | } |
98 | |
99 | void llvm_thread_join_impl(pthread_t Thread) { |
100 | int errnum; |
101 | |
102 | if ((errnum = ::pthread_join(th: Thread, thread_return: nullptr)) != 0) { |
103 | ReportErrnumFatal(Msg: "pthread_join failed" , errnum); |
104 | } |
105 | } |
106 | |
107 | pthread_t llvm_thread_get_id_impl(pthread_t Thread) { return Thread; } |
108 | |
109 | pthread_t llvm_thread_get_current_id_impl() { return ::pthread_self(); } |
110 | |
111 | } // namespace llvm |
112 | |
113 | uint64_t llvm::get_threadid() { |
114 | #if defined(__APPLE__) |
115 | // Calling "mach_thread_self()" bumps the reference count on the thread |
116 | // port, so we need to deallocate it. mach_task_self() doesn't bump the ref |
117 | // count. |
118 | static thread_local thread_port_t Self = [] { |
119 | thread_port_t InitSelf = mach_thread_self(); |
120 | mach_port_deallocate(mach_task_self(), Self); |
121 | return InitSelf; |
122 | }(); |
123 | return Self; |
124 | #elif defined(__FreeBSD__) |
125 | return uint64_t(pthread_getthreadid_np()); |
126 | #elif defined(__NetBSD__) |
127 | return uint64_t(_lwp_self()); |
128 | #elif defined(__OpenBSD__) |
129 | return uint64_t(getthrid()); |
130 | #elif defined(__ANDROID__) |
131 | return uint64_t(gettid()); |
132 | #elif defined(__linux__) |
133 | return uint64_t(syscall(SYS_gettid)); |
134 | #else |
135 | return uint64_t(pthread_self()); |
136 | #endif |
137 | } |
138 | |
139 | static constexpr uint32_t get_max_thread_name_length_impl() { |
140 | #if defined(__NetBSD__) |
141 | return PTHREAD_MAX_NAMELEN_NP; |
142 | #elif defined(__APPLE__) |
143 | return 64; |
144 | #elif defined(__linux__) |
145 | #if HAVE_PTHREAD_SETNAME_NP |
146 | return 16; |
147 | #else |
148 | return 0; |
149 | #endif |
150 | #elif defined(__FreeBSD__) || defined(__FreeBSD_kernel__) |
151 | return 16; |
152 | #elif defined(__OpenBSD__) |
153 | return 32; |
154 | #else |
155 | return 0; |
156 | #endif |
157 | } |
158 | |
159 | uint32_t llvm::get_max_thread_name_length() { |
160 | return get_max_thread_name_length_impl(); |
161 | } |
162 | |
163 | void llvm::set_thread_name(const Twine &Name) { |
164 | // Make sure the input is null terminated. |
165 | SmallString<64> Storage; |
166 | StringRef NameStr = Name.toNullTerminatedStringRef(Out&: Storage); |
167 | |
168 | // Truncate from the beginning, not the end, if the specified name is too |
169 | // long. For one, this ensures that the resulting string is still null |
170 | // terminated, but additionally the end of a long thread name will usually |
171 | // be more unique than the beginning, since a common pattern is for similar |
172 | // threads to share a common prefix. |
173 | // Note that the name length includes the null terminator. |
174 | if (get_max_thread_name_length() > 0) |
175 | NameStr = NameStr.take_back(N: get_max_thread_name_length() - 1); |
176 | (void)NameStr; |
177 | #if defined(__linux__) |
178 | #if (defined(__GLIBC__) && defined(_GNU_SOURCE)) || defined(__ANDROID__) |
179 | #if HAVE_PTHREAD_SETNAME_NP |
180 | ::pthread_setname_np(target_thread: ::pthread_self(), name: NameStr.data()); |
181 | #endif |
182 | #endif |
183 | #elif defined(__FreeBSD__) || defined(__OpenBSD__) |
184 | ::pthread_set_name_np(::pthread_self(), NameStr.data()); |
185 | #elif defined(__NetBSD__) |
186 | ::pthread_setname_np(::pthread_self(), "%s" , |
187 | const_cast<char *>(NameStr.data())); |
188 | #elif defined(__APPLE__) |
189 | ::pthread_setname_np(NameStr.data()); |
190 | #endif |
191 | } |
192 | |
193 | void llvm::get_thread_name(SmallVectorImpl<char> &Name) { |
194 | Name.clear(); |
195 | |
196 | #if defined(__FreeBSD__) || defined(__FreeBSD_kernel__) |
197 | int pid = ::getpid(); |
198 | uint64_t tid = get_threadid(); |
199 | |
200 | struct kinfo_proc *kp = nullptr, *nkp; |
201 | size_t len = 0; |
202 | int error; |
203 | int ctl[4] = {CTL_KERN, KERN_PROC, KERN_PROC_PID | KERN_PROC_INC_THREAD, |
204 | (int)pid}; |
205 | |
206 | while (1) { |
207 | error = sysctl(ctl, 4, kp, &len, nullptr, 0); |
208 | if (kp == nullptr || (error != 0 && errno == ENOMEM)) { |
209 | // Add extra space in case threads are added before next call. |
210 | len += sizeof(*kp) + len / 10; |
211 | nkp = (struct kinfo_proc *)::realloc(kp, len); |
212 | if (nkp == nullptr) { |
213 | free(kp); |
214 | return; |
215 | } |
216 | kp = nkp; |
217 | continue; |
218 | } |
219 | if (error != 0) |
220 | len = 0; |
221 | break; |
222 | } |
223 | |
224 | for (size_t i = 0; i < len / sizeof(*kp); i++) { |
225 | if (kp[i].ki_tid == (lwpid_t)tid) { |
226 | Name.append(kp[i].ki_tdname, kp[i].ki_tdname + strlen(kp[i].ki_tdname)); |
227 | break; |
228 | } |
229 | } |
230 | free(kp); |
231 | return; |
232 | #elif defined(__NetBSD__) |
233 | constexpr uint32_t len = get_max_thread_name_length_impl(); |
234 | char buf[len]; |
235 | ::pthread_getname_np(::pthread_self(), buf, len); |
236 | |
237 | Name.append(buf, buf + strlen(buf)); |
238 | #elif defined(__OpenBSD__) |
239 | constexpr uint32_t len = get_max_thread_name_length_impl(); |
240 | char buf[len]; |
241 | ::pthread_get_name_np(::pthread_self(), buf, len); |
242 | |
243 | Name.append(buf, buf + strlen(buf)); |
244 | #elif defined(__linux__) |
245 | #if HAVE_PTHREAD_GETNAME_NP |
246 | constexpr uint32_t len = get_max_thread_name_length_impl(); |
247 | char Buffer[len] = {'\0'}; // FIXME: working around MSan false positive. |
248 | if (0 == ::pthread_getname_np(target_thread: ::pthread_self(), buf: Buffer, buflen: len)) |
249 | Name.append(in_start: Buffer, in_end: Buffer + strlen(s: Buffer)); |
250 | #endif |
251 | #endif |
252 | } |
253 | |
254 | SetThreadPriorityResult llvm::set_thread_priority(ThreadPriority Priority) { |
255 | #if defined(__linux__) && defined(SCHED_IDLE) |
256 | // Some *really* old glibcs are missing SCHED_IDLE. |
257 | // http://man7.org/linux/man-pages/man3/pthread_setschedparam.3.html |
258 | // http://man7.org/linux/man-pages/man2/sched_setscheduler.2.html |
259 | sched_param priority; |
260 | // For each of the above policies, param->sched_priority must be 0. |
261 | priority.sched_priority = 0; |
262 | // SCHED_IDLE for running very low priority background jobs. |
263 | // SCHED_OTHER the standard round-robin time-sharing policy; |
264 | return !pthread_setschedparam( |
265 | target_thread: pthread_self(), |
266 | // FIXME: consider SCHED_BATCH for Low |
267 | policy: Priority == ThreadPriority::Default ? SCHED_OTHER : SCHED_IDLE, |
268 | param: &priority) |
269 | ? SetThreadPriorityResult::SUCCESS |
270 | : SetThreadPriorityResult::FAILURE; |
271 | #elif defined(__APPLE__) |
272 | // https://developer.apple.com/documentation/apple-silicon/tuning-your-code-s-performance-for-apple-silicon |
273 | // |
274 | // Background - Applies to work that isn’t visible to the user and may take |
275 | // significant time to complete. Examples include indexing, backing up, or |
276 | // synchronizing data. This class emphasizes energy efficiency. |
277 | // |
278 | // Utility - Applies to work that takes anywhere from a few seconds to a few |
279 | // minutes to complete. Examples include downloading a document or importing |
280 | // data. This class offers a balance between responsiveness, performance, and |
281 | // energy efficiency. |
282 | const auto qosClass = [&]() { |
283 | switch (Priority) { |
284 | case ThreadPriority::Background: |
285 | return QOS_CLASS_BACKGROUND; |
286 | case ThreadPriority::Low: |
287 | return QOS_CLASS_UTILITY; |
288 | case ThreadPriority::Default: |
289 | return QOS_CLASS_DEFAULT; |
290 | } |
291 | }(); |
292 | return !pthread_set_qos_class_self_np(qosClass, 0) |
293 | ? SetThreadPriorityResult::SUCCESS |
294 | : SetThreadPriorityResult::FAILURE; |
295 | #endif |
296 | return SetThreadPriorityResult::FAILURE; |
297 | } |
298 | |
299 | #include <thread> |
300 | |
301 | static int computeHostNumHardwareThreads() { |
302 | #if defined(__FreeBSD__) |
303 | cpuset_t mask; |
304 | CPU_ZERO(&mask); |
305 | if (cpuset_getaffinity(CPU_LEVEL_WHICH, CPU_WHICH_TID, -1, sizeof(mask), |
306 | &mask) == 0) |
307 | return CPU_COUNT(&mask); |
308 | #elif defined(__linux__) |
309 | cpu_set_t Set; |
310 | if (sched_getaffinity(pid: 0, cpusetsize: sizeof(Set), cpuset: &Set) == 0) |
311 | return CPU_COUNT(&Set); |
312 | #endif |
313 | // Guard against std::thread::hardware_concurrency() returning 0. |
314 | if (unsigned Val = std::thread::hardware_concurrency()) |
315 | return Val; |
316 | return 1; |
317 | } |
318 | |
319 | void llvm::ThreadPoolStrategy::apply_thread_strategy( |
320 | unsigned ThreadPoolNum) const {} |
321 | |
322 | llvm::BitVector llvm::get_thread_affinity_mask() { |
323 | // FIXME: Implement |
324 | llvm_unreachable("Not implemented!" ); |
325 | } |
326 | |
327 | unsigned llvm::get_cpus() { return 1; } |
328 | |
329 | #if defined(__linux__) && (defined(__i386__) || defined(__x86_64__)) |
330 | // On Linux, the number of physical cores can be computed from /proc/cpuinfo, |
331 | // using the number of unique physical/core id pairs. The following |
332 | // implementation reads the /proc/cpuinfo format on an x86_64 system. |
333 | static int computeHostNumPhysicalCores() { |
334 | // Enabled represents the number of physical id/core id pairs with at least |
335 | // one processor id enabled by the CPU affinity mask. |
336 | cpu_set_t Affinity, Enabled; |
337 | if (sched_getaffinity(pid: 0, cpusetsize: sizeof(Affinity), cpuset: &Affinity) != 0) |
338 | return -1; |
339 | CPU_ZERO(&Enabled); |
340 | |
341 | // Read /proc/cpuinfo as a stream (until EOF reached). It cannot be |
342 | // mmapped because it appears to have 0 size. |
343 | llvm::ErrorOr<std::unique_ptr<llvm::MemoryBuffer>> Text = |
344 | llvm::MemoryBuffer::getFileAsStream(Filename: "/proc/cpuinfo" ); |
345 | if (std::error_code EC = Text.getError()) { |
346 | llvm::errs() << "Can't read " |
347 | << "/proc/cpuinfo: " << EC.message() << "\n" ; |
348 | return -1; |
349 | } |
350 | SmallVector<StringRef, 8> strs; |
351 | (*Text)->getBuffer().split(A&: strs, Separator: "\n" , /*MaxSplit=*/MaxSplit: -1, |
352 | /*KeepEmpty=*/KeepEmpty: false); |
353 | int CurProcessor = -1; |
354 | int CurPhysicalId = -1; |
355 | int CurSiblings = -1; |
356 | int CurCoreId = -1; |
357 | for (StringRef Line : strs) { |
358 | std::pair<StringRef, StringRef> Data = Line.split(Separator: ':'); |
359 | auto Name = Data.first.trim(); |
360 | auto Val = Data.second.trim(); |
361 | // These fields are available if the kernel is configured with CONFIG_SMP. |
362 | if (Name == "processor" ) |
363 | Val.getAsInteger(Radix: 10, Result&: CurProcessor); |
364 | else if (Name == "physical id" ) |
365 | Val.getAsInteger(Radix: 10, Result&: CurPhysicalId); |
366 | else if (Name == "siblings" ) |
367 | Val.getAsInteger(Radix: 10, Result&: CurSiblings); |
368 | else if (Name == "core id" ) { |
369 | Val.getAsInteger(Radix: 10, Result&: CurCoreId); |
370 | // The processor id corresponds to an index into cpu_set_t. |
371 | if (CPU_ISSET(CurProcessor, &Affinity)) |
372 | CPU_SET(CurPhysicalId * CurSiblings + CurCoreId, &Enabled); |
373 | } |
374 | } |
375 | return CPU_COUNT(&Enabled); |
376 | } |
377 | #elif (defined(__linux__) && defined(__s390x__)) || defined(_AIX) |
378 | static int computeHostNumPhysicalCores() { |
379 | return sysconf(_SC_NPROCESSORS_ONLN); |
380 | } |
381 | #elif defined(__linux__) && !defined(__ANDROID__) |
382 | static int computeHostNumPhysicalCores() { |
383 | cpu_set_t Affinity; |
384 | if (sched_getaffinity(0, sizeof(Affinity), &Affinity) == 0) |
385 | return CPU_COUNT(&Affinity); |
386 | |
387 | // The call to sched_getaffinity() may have failed because the Affinity |
388 | // mask is too small for the number of CPU's on the system (i.e. the |
389 | // system has more than 1024 CPUs). Allocate a mask large enough for |
390 | // twice as many CPUs. |
391 | cpu_set_t *DynAffinity; |
392 | DynAffinity = CPU_ALLOC(2048); |
393 | if (sched_getaffinity(0, CPU_ALLOC_SIZE(2048), DynAffinity) == 0) { |
394 | int NumCPUs = CPU_COUNT(DynAffinity); |
395 | CPU_FREE(DynAffinity); |
396 | return NumCPUs; |
397 | } |
398 | return -1; |
399 | } |
400 | #elif defined(__APPLE__) |
401 | // Gets the number of *physical cores* on the machine. |
402 | static int computeHostNumPhysicalCores() { |
403 | uint32_t count; |
404 | size_t len = sizeof(count); |
405 | sysctlbyname("hw.physicalcpu" , &count, &len, NULL, 0); |
406 | if (count < 1) { |
407 | int nm[2]; |
408 | nm[0] = CTL_HW; |
409 | nm[1] = HW_AVAILCPU; |
410 | sysctl(nm, 2, &count, &len, NULL, 0); |
411 | if (count < 1) |
412 | return -1; |
413 | } |
414 | return count; |
415 | } |
416 | #elif defined(__MVS__) |
417 | static int computeHostNumPhysicalCores() { |
418 | enum { |
419 | // Byte offset of the pointer to the Communications Vector Table (CVT) in |
420 | // the Prefixed Save Area (PSA). The table entry is a 31-bit pointer and |
421 | // will be zero-extended to uintptr_t. |
422 | FLCCVT = 16, |
423 | // Byte offset of the pointer to the Common System Data Area (CSD) in the |
424 | // CVT. The table entry is a 31-bit pointer and will be zero-extended to |
425 | // uintptr_t. |
426 | CVTCSD = 660, |
427 | // Byte offset to the number of live CPs in the LPAR, stored as a signed |
428 | // 32-bit value in the table. |
429 | CSD_NUMBER_ONLINE_STANDARD_CPS = 264, |
430 | }; |
431 | char *PSA = 0; |
432 | char *CVT = reinterpret_cast<char *>( |
433 | static_cast<uintptr_t>(reinterpret_cast<unsigned int &>(PSA[FLCCVT]))); |
434 | char *CSD = reinterpret_cast<char *>( |
435 | static_cast<uintptr_t>(reinterpret_cast<unsigned int &>(CVT[CVTCSD]))); |
436 | return reinterpret_cast<int &>(CSD[CSD_NUMBER_ONLINE_STANDARD_CPS]); |
437 | } |
438 | #else |
439 | // On other systems, return -1 to indicate unknown. |
440 | static int computeHostNumPhysicalCores() { return -1; } |
441 | #endif |
442 | |
443 | int llvm::get_physical_cores() { |
444 | static int NumCores = computeHostNumPhysicalCores(); |
445 | return NumCores; |
446 | } |
447 | |