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 #include <windows.h>
54 #endif
55
56 #ifdef __FreeBSD__
57 /* pthread_np.h -> sys/param.h -> machine/param.h
58 * - defines ALIGN which clashes with our ALIGN
59 */
60 #undef ALIGN
61 #define cpu_set_t cpuset_t
62 #endif
63
64 /* For util_set_thread_affinity to size the mask. */
65 #define UTIL_MAX_CPUS 1024 /* this should be enough */
66 #define UTIL_MAX_L3_CACHES UTIL_MAX_CPUS
67
68 /* Some highly performance-sensitive thread-local variables like the current GL
69 * context are declared with the initial-exec model on Linux. glibc allocates a
70 * fixed number of extra slots for initial-exec TLS variables at startup, and
71 * Mesa relies on (even if it's dlopen()ed after init) being able to fit into
72 * those. This model saves the call to look up the address of the TLS variable.
73 *
74 * However, if we don't have this TLS model available on the platform, then we
75 * still want to use normal TLS (which involves a function call, but not the
76 * expensive pthread_getspecific() or its equivalent).
77 */
78 #ifdef _MSC_VER
79 #define __THREAD_INITIAL_EXEC __declspec(thread)
80 #elif defined(ANDROID)
81 /* Android 29 gained ELF TLS support, but it doesn't support initial-exec and
82 * it will throw:
83 *
84 * dlopen failed: TLS symbol "(null)" in dlopened
85 * "/vendor/lib64/egl/libEGL_mesa.so" referenced from
86 * "/vendor/lib64/egl/libEGL_mesa.so" using IE access model.
87 */
88 #define __THREAD_INITIAL_EXEC __thread
89 #else
90 #define __THREAD_INITIAL_EXEC __thread __attribute__((tls_model("initial-exec")))
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(int (* routine)(void *),void * param)107 static inline thrd_t u_thread_create(int (*routine)(void *), void *param)
108 {
109 thrd_t thread;
110 #ifdef HAVE_PTHREAD
111 sigset_t saved_set, new_set;
112 int ret;
113
114 sigfillset(&new_set);
115 sigdelset(&new_set, SIGSYS);
116 pthread_sigmask(SIG_BLOCK, &new_set, &saved_set);
117 ret = thrd_create( &thread, routine, param );
118 pthread_sigmask(SIG_SETMASK, &saved_set, NULL);
119 #else
120 int ret;
121 ret = thrd_create( &thread, routine, param );
122 #endif
123 if (ret)
124 return 0;
125
126 return thread;
127 }
128
u_thread_setname(const char * name)129 static inline void u_thread_setname( const char *name )
130 {
131 #if defined(HAVE_PTHREAD)
132 #if DETECT_OS_LINUX || DETECT_OS_CYGWIN || DETECT_OS_SOLARIS
133 int ret = pthread_setname_np(pthread_self(), name);
134 if (ret == ERANGE) {
135 char buf[16];
136 const size_t len = MIN2(strlen(name), ARRAY_SIZE(buf) - 1);
137 memcpy(buf, name, len);
138 buf[len] = '\0';
139 pthread_setname_np(pthread_self(), buf);
140 }
141 #elif DETECT_OS_FREEBSD || DETECT_OS_OPENBSD
142 pthread_set_name_np(pthread_self(), name);
143 #elif DETECT_OS_NETBSD
144 pthread_setname_np(pthread_self(), "%s", (void *)name);
145 #elif DETECT_OS_APPLE
146 pthread_setname_np(name);
147 #elif DETECT_OS_HAIKU
148 rename_thread(find_thread(NULL), name);
149 #else
150 #warning Not sure how to call pthread_setname_np
151 #endif
152 #endif
153 (void)name;
154 }
155
156 /**
157 * Set thread affinity.
158 *
159 * \param thread Thread
160 * \param mask Set this affinity mask
161 * \param old_mask Previous affinity mask returned if not NULL
162 * \param num_mask_bits Number of bits in both masks
163 * \return true on success
164 */
165 static inline bool
util_set_thread_affinity(thrd_t thread,const uint32_t * mask,uint32_t * old_mask,unsigned num_mask_bits)166 util_set_thread_affinity(thrd_t thread,
167 const uint32_t *mask,
168 uint32_t *old_mask,
169 unsigned num_mask_bits)
170 {
171 #if defined(HAVE_PTHREAD_SETAFFINITY)
172 cpu_set_t cpuset;
173
174 if (old_mask) {
175 if (pthread_getaffinity_np(thread, sizeof(cpuset), &cpuset) != 0)
176 return false;
177
178 memset(old_mask, 0, num_mask_bits / 8);
179 for (unsigned i = 0; i < num_mask_bits && i < CPU_SETSIZE; i++) {
180 if (CPU_ISSET(i, &cpuset))
181 old_mask[i / 32] |= 1u << (i % 32);
182 }
183 }
184
185 CPU_ZERO(&cpuset);
186 for (unsigned i = 0; i < num_mask_bits && i < CPU_SETSIZE; i++) {
187 if (mask[i / 32] & (1u << (i % 32)))
188 CPU_SET(i, &cpuset);
189 }
190 return pthread_setaffinity_np(thread, sizeof(cpuset), &cpuset) == 0;
191
192 #elif defined(_WIN32) && !defined(__CYGWIN__)
193 DWORD_PTR m = mask[0];
194
195 if (sizeof(m) > 4 && num_mask_bits > 32)
196 m |= (uint64_t)mask[1] << 32;
197
198 m = SetThreadAffinityMask(thread, m);
199 if (!m)
200 return false;
201
202 if (old_mask) {
203 memset(old_mask, 0, num_mask_bits / 8);
204
205 old_mask[0] = m;
206 #ifdef _WIN64
207 old_mask[1] = m >> 32;
208 #endif
209 }
210
211 return true;
212 #else
213 return false;
214 #endif
215 }
216
217 static inline bool
util_set_current_thread_affinity(const uint32_t * mask,uint32_t * old_mask,unsigned num_mask_bits)218 util_set_current_thread_affinity(const uint32_t *mask,
219 uint32_t *old_mask,
220 unsigned num_mask_bits)
221 {
222 #if defined(HAVE_PTHREAD_SETAFFINITY)
223 return util_set_thread_affinity(pthread_self(), mask, old_mask,
224 num_mask_bits);
225
226 #elif defined(_WIN32) && !defined(__CYGWIN__)
227 /* The GetCurrentThreadId() handle is only valid within the current thread. */
228 return util_set_thread_affinity(GetCurrentThread(), mask, old_mask,
229 num_mask_bits);
230
231 #else
232 return false;
233 #endif
234 }
235
236
237 /*
238 * Thread statistics.
239 */
240
241 /* Return the time of a thread's CPU time clock. */
242 static inline int64_t
util_thread_get_time_nano(thrd_t thread)243 util_thread_get_time_nano(thrd_t thread)
244 {
245 #if defined(HAVE_PTHREAD) && !defined(__APPLE__) && !defined(__HAIKU__)
246 struct timespec ts;
247 clockid_t cid;
248
249 pthread_getcpuclockid(thread, &cid);
250 clock_gettime(cid, &ts);
251 return (int64_t)ts.tv_sec * 1000000000 + ts.tv_nsec;
252 #else
253 return 0;
254 #endif
255 }
256
257 /* Return the time of the current thread's CPU time clock. */
258 static inline int64_t
util_current_thread_get_time_nano(void)259 util_current_thread_get_time_nano(void)
260 {
261 #if defined(HAVE_PTHREAD)
262 return util_thread_get_time_nano(pthread_self());
263
264 #elif defined(_WIN32) && !defined(__CYGWIN__)
265 /* The GetCurrentThreadId() handle is only valid within the current thread. */
266 return util_thread_get_time_nano(GetCurrentThread());
267
268 #else
269 return 0;
270 #endif
271 }
272
u_thread_is_self(thrd_t thread)273 static inline bool u_thread_is_self(thrd_t thread)
274 {
275 #if defined(HAVE_PTHREAD)
276 return pthread_equal(pthread_self(), thread);
277 #endif
278 return false;
279 }
280
281 /*
282 * util_barrier
283 */
284
285 #if defined(HAVE_PTHREAD) && !defined(__APPLE__) && !defined(__HAIKU__)
286
287 typedef pthread_barrier_t util_barrier;
288
util_barrier_init(util_barrier * barrier,unsigned count)289 static inline void util_barrier_init(util_barrier *barrier, unsigned count)
290 {
291 pthread_barrier_init(barrier, NULL, count);
292 }
293
util_barrier_destroy(util_barrier * barrier)294 static inline void util_barrier_destroy(util_barrier *barrier)
295 {
296 pthread_barrier_destroy(barrier);
297 }
298
util_barrier_wait(util_barrier * barrier)299 static inline void util_barrier_wait(util_barrier *barrier)
300 {
301 pthread_barrier_wait(barrier);
302 }
303
304
305 #else /* If the OS doesn't have its own, implement barriers using a mutex and a condvar */
306
307 typedef struct {
308 unsigned count;
309 unsigned waiters;
310 uint64_t sequence;
311 mtx_t mutex;
312 cnd_t condvar;
313 } util_barrier;
314
util_barrier_init(util_barrier * barrier,unsigned count)315 static inline void util_barrier_init(util_barrier *barrier, unsigned count)
316 {
317 barrier->count = count;
318 barrier->waiters = 0;
319 barrier->sequence = 0;
320 (void) mtx_init(&barrier->mutex, mtx_plain);
321 cnd_init(&barrier->condvar);
322 }
323
util_barrier_destroy(util_barrier * barrier)324 static inline void util_barrier_destroy(util_barrier *barrier)
325 {
326 assert(barrier->waiters == 0);
327 mtx_destroy(&barrier->mutex);
328 cnd_destroy(&barrier->condvar);
329 }
330
util_barrier_wait(util_barrier * barrier)331 static inline void util_barrier_wait(util_barrier *barrier)
332 {
333 mtx_lock(&barrier->mutex);
334
335 assert(barrier->waiters < barrier->count);
336 barrier->waiters++;
337
338 if (barrier->waiters < barrier->count) {
339 uint64_t sequence = barrier->sequence;
340
341 do {
342 cnd_wait(&barrier->condvar, &barrier->mutex);
343 } while (sequence == barrier->sequence);
344 } else {
345 barrier->waiters = 0;
346 barrier->sequence++;
347 cnd_broadcast(&barrier->condvar);
348 }
349
350 mtx_unlock(&barrier->mutex);
351 }
352
353 #endif
354
355 /*
356 * Thread-id's.
357 *
358 * thrd_current() is not portable to windows (or at least not in a desirable
359 * way), so thread_id's provide an alternative mechanism
360 */
361
362 #ifdef _WIN32
363 typedef DWORD thread_id;
364 #else
365 typedef thrd_t thread_id;
366 #endif
367
368 static inline thread_id
util_get_thread_id(void)369 util_get_thread_id(void)
370 {
371 /*
372 * XXX: Callers of of this function assume it is a lightweight function.
373 * But unfortunately C11's thrd_current() gives no such guarantees. In
374 * fact, it's pretty hard to have a compliant implementation of
375 * thrd_current() on Windows with such characteristics. So for now, we
376 * side-step this mess and use Windows thread primitives directly here.
377 */
378 #ifdef _WIN32
379 return GetCurrentThreadId();
380 #else
381 return thrd_current();
382 #endif
383 }
384
385
386 static inline int
util_thread_id_equal(thread_id t1,thread_id t2)387 util_thread_id_equal(thread_id t1, thread_id t2)
388 {
389 #ifdef _WIN32
390 return t1 == t2;
391 #else
392 return thrd_equal(t1, t2);
393 #endif
394 }
395
396 #endif /* U_THREAD_H_ */
397