1 /**************************************************************************
2 *
3 * Copyright 1999-2006 Brian Paul
4 * Copyright 2008 VMware, Inc.
5 * All Rights Reserved.
6 *
7 * Permission is hereby granted, free of charge, to any person obtaining a
8 * copy of this software and associated documentation files (the "Software"),
9 * to deal in the Software without restriction, including without limitation
10 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
11 * and/or sell copies of the Software, and to permit persons to whom the
12 * Software is furnished to do so, subject to the following conditions:
13 *
14 * The above copyright notice and this permission notice shall be included
15 * in all copies or substantial portions of the Software.
16 *
17 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
18 * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
19 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
20 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR
21 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
22 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
23 * OTHER DEALINGS IN THE SOFTWARE.
24 *
25 **************************************************************************/
26
27 #ifndef U_THREAD_H_
28 #define U_THREAD_H_
29
30 #include <errno.h>
31 #include <stdint.h>
32 #include <stdbool.h>
33 #include <string.h>
34
35 #include "c11/threads.h"
36 #include "detect_os.h"
37 #include "macros.h"
38
39 #ifdef HAVE_PTHREAD
40 #include <signal.h>
41 #ifdef HAVE_PTHREAD_NP_H
42 #include <pthread_np.h>
43 #endif
44 #endif
45
46 #ifdef __HAIKU__
47 #include <OS.h>
48 #endif
49
50 #if DETECT_OS_LINUX && !defined(ANDROID)
51 #include <sched.h>
52 #elif defined(_WIN32) && !defined(__CYGWIN__) && _WIN32_WINNT >= 0x0600
53 #ifndef WIN32_LEAN_AND_MEAN
54 #define WIN32_LEAN_AND_MEAN 1
55 #endif
56 #include <windows.h>
57 #endif
58
59 #ifdef __FreeBSD__
60 /* pthread_np.h -> sys/param.h -> machine/param.h
61 * - defines ALIGN which clashes with our ALIGN
62 */
63 #undef ALIGN
64 #define cpu_set_t cpuset_t
65 #endif
66
67 /* For util_set_thread_affinity to size the mask. */
68 #define UTIL_MAX_CPUS 1024 /* this should be enough */
69 #define UTIL_MAX_L3_CACHES UTIL_MAX_CPUS
70
71 /* Some highly performance-sensitive thread-local variables like the current GL
72 * context are declared with the initial-exec model on Linux. glibc allocates a
73 * fixed number of extra slots for initial-exec TLS variables at startup, and
74 * Mesa relies on (even if it's dlopen()ed after init) being able to fit into
75 * those. This model saves the call to look up the address of the TLS variable.
76 *
77 * However, if we don't have this TLS model available on the platform, then we
78 * still want to use normal TLS (which involves a function call, but not the
79 * expensive pthread_getspecific() or its equivalent).
80 */
81 #if DETECT_OS_APPLE
82 /* Apple Clang emits wrappers when using thread_local that break module linkage,
83 * but not with __thread
84 */
85 #define __THREAD_INITIAL_EXEC __thread
86 #elif defined(__GLIBC__)
87 #define __THREAD_INITIAL_EXEC thread_local __attribute__((tls_model("initial-exec")))
88 #define REALLY_INITIAL_EXEC
89 #else
90 #define __THREAD_INITIAL_EXEC thread_local
91 #endif
92
93 static inline int
util_get_current_cpu(void)94 util_get_current_cpu(void)
95 {
96 #if DETECT_OS_LINUX && !defined(ANDROID)
97 return sched_getcpu();
98
99 #elif defined(_WIN32) && !defined(__CYGWIN__) && _WIN32_WINNT >= 0x0600
100 return GetCurrentProcessorNumber();
101
102 #else
103 return -1;
104 #endif
105 }
106
u_thread_create(thrd_t * thrd,int (* routine)(void *),void * param)107 static inline int u_thread_create(thrd_t *thrd, int (*routine)(void *), void *param)
108 {
109 int ret = thrd_error;
110 #ifdef HAVE_PTHREAD
111 sigset_t saved_set, new_set;
112
113 sigfillset(&new_set);
114 sigdelset(&new_set, SIGSYS);
115
116 /* SIGSEGV is commonly used by Vulkan API tracing layers in order to track
117 * accesses in device memory mapped to user space. Blocking the signal hinders
118 * that tracking mechanism.
119 */
120 sigdelset(&new_set, SIGSEGV);
121 pthread_sigmask(SIG_BLOCK, &new_set, &saved_set);
122 ret = thrd_create(thrd, routine, param);
123 pthread_sigmask(SIG_SETMASK, &saved_set, NULL);
124 #else
125 ret = thrd_create(thrd, routine, param);
126 #endif
127
128 return ret;
129 }
130
u_thread_setname(const char * name)131 static inline void u_thread_setname( const char *name )
132 {
133 #if defined(HAVE_PTHREAD)
134 #if DETECT_OS_LINUX || DETECT_OS_CYGWIN || DETECT_OS_SOLARIS
135 int ret = pthread_setname_np(pthread_self(), name);
136 if (ret == ERANGE) {
137 char buf[16];
138 const size_t len = MIN2(strlen(name), ARRAY_SIZE(buf) - 1);
139 memcpy(buf, name, len);
140 buf[len] = '\0';
141 pthread_setname_np(pthread_self(), buf);
142 }
143 #elif DETECT_OS_FREEBSD || DETECT_OS_OPENBSD
144 pthread_set_name_np(pthread_self(), name);
145 #elif DETECT_OS_NETBSD
146 pthread_setname_np(pthread_self(), "%s", (void *)name);
147 #elif DETECT_OS_APPLE
148 pthread_setname_np(name);
149 #elif DETECT_OS_HAIKU
150 rename_thread(find_thread(NULL), name);
151 #else
152 #warning Not sure how to call pthread_setname_np
153 #endif
154 #endif
155 (void)name;
156 }
157
158 /**
159 * Set thread affinity.
160 *
161 * \param thread Thread
162 * \param mask Set this affinity mask
163 * \param old_mask Previous affinity mask returned if not NULL
164 * \param num_mask_bits Number of bits in both masks
165 * \return true on success
166 */
167 static inline bool
util_set_thread_affinity(thrd_t thread,const uint32_t * mask,uint32_t * old_mask,unsigned num_mask_bits)168 util_set_thread_affinity(thrd_t thread,
169 const uint32_t *mask,
170 uint32_t *old_mask,
171 unsigned num_mask_bits)
172 {
173 #if defined(HAVE_PTHREAD_SETAFFINITY)
174 cpu_set_t cpuset;
175
176 if (old_mask) {
177 if (pthread_getaffinity_np(thread, sizeof(cpuset), &cpuset) != 0)
178 return false;
179
180 memset(old_mask, 0, num_mask_bits / 8);
181 for (unsigned i = 0; i < num_mask_bits && i < CPU_SETSIZE; i++) {
182 if (CPU_ISSET(i, &cpuset))
183 old_mask[i / 32] |= 1u << (i % 32);
184 }
185 }
186
187 CPU_ZERO(&cpuset);
188 for (unsigned i = 0; i < num_mask_bits && i < CPU_SETSIZE; i++) {
189 if (mask[i / 32] & (1u << (i % 32)))
190 CPU_SET(i, &cpuset);
191 }
192 return pthread_setaffinity_np(thread, sizeof(cpuset), &cpuset) == 0;
193
194 #elif defined(_WIN32) && !defined(__CYGWIN__)
195 DWORD_PTR m = mask[0];
196
197 if (sizeof(m) > 4 && num_mask_bits > 32)
198 m |= (uint64_t)mask[1] << 32;
199
200 m = SetThreadAffinityMask(thread, m);
201 if (!m)
202 return false;
203
204 if (old_mask) {
205 memset(old_mask, 0, num_mask_bits / 8);
206
207 old_mask[0] = m;
208 #ifdef _WIN64
209 old_mask[1] = m >> 32;
210 #endif
211 }
212
213 return true;
214 #else
215 return false;
216 #endif
217 }
218
219 static inline bool
util_set_current_thread_affinity(const uint32_t * mask,uint32_t * old_mask,unsigned num_mask_bits)220 util_set_current_thread_affinity(const uint32_t *mask,
221 uint32_t *old_mask,
222 unsigned num_mask_bits)
223 {
224 #if defined(HAVE_PTHREAD_SETAFFINITY)
225 return util_set_thread_affinity(pthread_self(), mask, old_mask,
226 num_mask_bits);
227
228 #elif defined(_WIN32) && !defined(__CYGWIN__)
229 /* The GetCurrentThreadId() handle is only valid within the current thread. */
230 return util_set_thread_affinity(GetCurrentThread(), mask, old_mask,
231 num_mask_bits);
232
233 #else
234 return false;
235 #endif
236 }
237
238
239 /*
240 * Thread statistics.
241 */
242
243 /* Return the time of a thread's CPU time clock. */
244 static inline int64_t
util_thread_get_time_nano(thrd_t thread)245 util_thread_get_time_nano(thrd_t thread)
246 {
247 #if defined(HAVE_PTHREAD) && !defined(__APPLE__) && !defined(__HAIKU__)
248 struct timespec ts;
249 clockid_t cid;
250
251 pthread_getcpuclockid(thread, &cid);
252 clock_gettime(cid, &ts);
253 return (int64_t)ts.tv_sec * 1000000000 + ts.tv_nsec;
254 #else
255 (void)thread;
256 return 0;
257 #endif
258 }
259
260 /* Return the time of the current thread's CPU time clock. */
261 static inline int64_t
util_current_thread_get_time_nano(void)262 util_current_thread_get_time_nano(void)
263 {
264 #if defined(HAVE_PTHREAD)
265 return util_thread_get_time_nano(pthread_self());
266
267 #elif defined(_WIN32) && !defined(__CYGWIN__)
268 /* The GetCurrentThreadId() handle is only valid within the current thread. */
269 return util_thread_get_time_nano(GetCurrentThread());
270
271 #else
272 return 0;
273 #endif
274 }
275
u_thread_is_self(thrd_t thread)276 static inline bool u_thread_is_self(thrd_t thread)
277 {
278 #if defined(HAVE_PTHREAD)
279 return pthread_equal(pthread_self(), thread);
280 #endif
281 return false;
282 }
283
284 /*
285 * util_barrier
286 */
287
288 #if defined(HAVE_PTHREAD) && !defined(__APPLE__) && !defined(__HAIKU__)
289
290 typedef pthread_barrier_t util_barrier;
291
util_barrier_init(util_barrier * barrier,unsigned count)292 static inline void util_barrier_init(util_barrier *barrier, unsigned count)
293 {
294 pthread_barrier_init(barrier, NULL, count);
295 }
296
util_barrier_destroy(util_barrier * barrier)297 static inline void util_barrier_destroy(util_barrier *barrier)
298 {
299 pthread_barrier_destroy(barrier);
300 }
301
util_barrier_wait(util_barrier * barrier)302 static inline bool util_barrier_wait(util_barrier *barrier)
303 {
304 return pthread_barrier_wait(barrier) == PTHREAD_BARRIER_SERIAL_THREAD;
305 }
306
307
308 #else /* If the OS doesn't have its own, implement barriers using a mutex and a condvar */
309
310 typedef struct {
311 unsigned count;
312 unsigned waiters;
313 uint64_t sequence;
314 mtx_t mutex;
315 cnd_t condvar;
316 } util_barrier;
317
util_barrier_init(util_barrier * barrier,unsigned count)318 static inline void util_barrier_init(util_barrier *barrier, unsigned count)
319 {
320 barrier->count = count;
321 barrier->waiters = 0;
322 barrier->sequence = 0;
323 (void) mtx_init(&barrier->mutex, mtx_plain);
324 cnd_init(&barrier->condvar);
325 }
326
util_barrier_destroy(util_barrier * barrier)327 static inline void util_barrier_destroy(util_barrier *barrier)
328 {
329 assert(barrier->waiters == 0);
330 mtx_destroy(&barrier->mutex);
331 cnd_destroy(&barrier->condvar);
332 }
333
util_barrier_wait(util_barrier * barrier)334 static inline bool util_barrier_wait(util_barrier *barrier)
335 {
336 mtx_lock(&barrier->mutex);
337
338 assert(barrier->waiters < barrier->count);
339 barrier->waiters++;
340
341 if (barrier->waiters < barrier->count) {
342 uint64_t sequence = barrier->sequence;
343
344 do {
345 cnd_wait(&barrier->condvar, &barrier->mutex);
346 } while (sequence == barrier->sequence);
347 } else {
348 barrier->waiters = 0;
349 barrier->sequence++;
350 cnd_broadcast(&barrier->condvar);
351 }
352
353 mtx_unlock(&barrier->mutex);
354
355 return true;
356 }
357
358 #endif
359
360 /*
361 * Thread-id's.
362 *
363 * thrd_current() is not portable to windows (or at least not in a desirable
364 * way), so thread_id's provide an alternative mechanism
365 */
366
367 #ifdef _WIN32
368 typedef DWORD thread_id;
369 #else
370 typedef thrd_t thread_id;
371 #endif
372
373 static inline thread_id
util_get_thread_id(void)374 util_get_thread_id(void)
375 {
376 /*
377 * XXX: Callers of of this function assume it is a lightweight function.
378 * But unfortunately C11's thrd_current() gives no such guarantees. In
379 * fact, it's pretty hard to have a compliant implementation of
380 * thrd_current() on Windows with such characteristics. So for now, we
381 * side-step this mess and use Windows thread primitives directly here.
382 */
383 #ifdef _WIN32
384 return GetCurrentThreadId();
385 #else
386 return thrd_current();
387 #endif
388 }
389
390
391 static inline int
util_thread_id_equal(thread_id t1,thread_id t2)392 util_thread_id_equal(thread_id t1, thread_id t2)
393 {
394 #ifdef _WIN32
395 return t1 == t2;
396 #else
397 return thrd_equal(t1, t2);
398 #endif
399 }
400
401 #endif /* U_THREAD_H_ */
402