Improve DLA_FMA for FMA4
[glibc.git] / nptl / nptl-init.c
blobdb45cab230e65a4b41fba3f1ece7186daa49df3b
1 /* Copyright (C) 2002-2007, 2008, 2009, 2011 Free Software Foundation, Inc.
2 This file is part of the GNU C Library.
3 Contributed by Ulrich Drepper <drepper@redhat.com>, 2002.
5 The GNU C Library is free software; you can redistribute it and/or
6 modify it under the terms of the GNU Lesser General Public
7 License as published by the Free Software Foundation; either
8 version 2.1 of the License, or (at your option) any later version.
10 The GNU C Library is distributed in the hope that it will be useful,
11 but WITHOUT ANY WARRANTY; without even the implied warranty of
12 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 Lesser General Public License for more details.
15 You should have received a copy of the GNU Lesser General Public
16 License along with the GNU C Library; if not, write to the Free
17 Software Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
18 02111-1307 USA. */
20 #include <assert.h>
21 #include <errno.h>
22 #include <limits.h>
23 #include <signal.h>
24 #include <stdlib.h>
25 #include <unistd.h>
26 #include <sys/param.h>
27 #include <sys/resource.h>
28 #include <pthreadP.h>
29 #include <atomic.h>
30 #include <ldsodefs.h>
31 #include <tls.h>
32 #include <list.h>
33 #include <fork.h>
34 #include <version.h>
35 #include <shlib-compat.h>
36 #include <smp.h>
37 #include <lowlevellock.h>
38 #include <kernel-features.h>
41 /* Size and alignment of static TLS block. */
42 size_t __static_tls_size;
43 size_t __static_tls_align_m1;
45 #ifndef __ASSUME_SET_ROBUST_LIST
46 /* Negative if we do not have the system call and we can use it. */
47 int __set_robust_list_avail;
48 # define set_robust_list_not_avail() \
49 __set_robust_list_avail = -1
50 #else
51 # define set_robust_list_not_avail() do { } while (0)
52 #endif
54 #ifndef __ASSUME_FUTEX_CLOCK_REALTIME
55 /* Nonzero if we do not have FUTEX_CLOCK_REALTIME. */
56 int __have_futex_clock_realtime;
57 # define __set_futex_clock_realtime() \
58 __have_futex_clock_realtime = 1
59 #else
60 #define __set_futex_clock_realtime() do { } while (0)
61 #endif
63 /* Version of the library, used in libthread_db to detect mismatches. */
64 static const char nptl_version[] __attribute_used__ = VERSION;
67 #ifndef SHARED
68 extern void __libc_setup_tls (size_t tcbsize, size_t tcbalign);
69 #endif
71 #ifdef SHARED
72 static
73 #else
74 extern
75 #endif
76 void __nptl_set_robust (struct pthread *);
78 #ifdef SHARED
79 static void nptl_freeres (void);
82 static const struct pthread_functions pthread_functions =
84 .ptr_pthread_attr_destroy = __pthread_attr_destroy,
85 # if SHLIB_COMPAT(libpthread, GLIBC_2_0, GLIBC_2_1)
86 .ptr___pthread_attr_init_2_0 = __pthread_attr_init_2_0,
87 # endif
88 .ptr___pthread_attr_init_2_1 = __pthread_attr_init_2_1,
89 .ptr_pthread_attr_getdetachstate = __pthread_attr_getdetachstate,
90 .ptr_pthread_attr_setdetachstate = __pthread_attr_setdetachstate,
91 .ptr_pthread_attr_getinheritsched = __pthread_attr_getinheritsched,
92 .ptr_pthread_attr_setinheritsched = __pthread_attr_setinheritsched,
93 .ptr_pthread_attr_getschedparam = __pthread_attr_getschedparam,
94 .ptr_pthread_attr_setschedparam = __pthread_attr_setschedparam,
95 .ptr_pthread_attr_getschedpolicy = __pthread_attr_getschedpolicy,
96 .ptr_pthread_attr_setschedpolicy = __pthread_attr_setschedpolicy,
97 .ptr_pthread_attr_getscope = __pthread_attr_getscope,
98 .ptr_pthread_attr_setscope = __pthread_attr_setscope,
99 .ptr_pthread_condattr_destroy = __pthread_condattr_destroy,
100 .ptr_pthread_condattr_init = __pthread_condattr_init,
101 .ptr___pthread_cond_broadcast = __pthread_cond_broadcast,
102 .ptr___pthread_cond_destroy = __pthread_cond_destroy,
103 .ptr___pthread_cond_init = __pthread_cond_init,
104 .ptr___pthread_cond_signal = __pthread_cond_signal,
105 .ptr___pthread_cond_wait = __pthread_cond_wait,
106 .ptr___pthread_cond_timedwait = __pthread_cond_timedwait,
107 # if SHLIB_COMPAT(libpthread, GLIBC_2_0, GLIBC_2_3_2)
108 .ptr___pthread_cond_broadcast_2_0 = __pthread_cond_broadcast_2_0,
109 .ptr___pthread_cond_destroy_2_0 = __pthread_cond_destroy_2_0,
110 .ptr___pthread_cond_init_2_0 = __pthread_cond_init_2_0,
111 .ptr___pthread_cond_signal_2_0 = __pthread_cond_signal_2_0,
112 .ptr___pthread_cond_wait_2_0 = __pthread_cond_wait_2_0,
113 .ptr___pthread_cond_timedwait_2_0 = __pthread_cond_timedwait_2_0,
114 # endif
115 .ptr_pthread_equal = __pthread_equal,
116 .ptr___pthread_exit = __pthread_exit,
117 .ptr_pthread_getschedparam = __pthread_getschedparam,
118 .ptr_pthread_setschedparam = __pthread_setschedparam,
119 .ptr_pthread_mutex_destroy = INTUSE(__pthread_mutex_destroy),
120 .ptr_pthread_mutex_init = INTUSE(__pthread_mutex_init),
121 .ptr_pthread_mutex_lock = INTUSE(__pthread_mutex_lock),
122 .ptr_pthread_mutex_unlock = INTUSE(__pthread_mutex_unlock),
123 .ptr_pthread_self = __pthread_self,
124 .ptr_pthread_setcancelstate = __pthread_setcancelstate,
125 .ptr_pthread_setcanceltype = __pthread_setcanceltype,
126 .ptr___pthread_cleanup_upto = __pthread_cleanup_upto,
127 .ptr___pthread_once = __pthread_once_internal,
128 .ptr___pthread_rwlock_rdlock = __pthread_rwlock_rdlock_internal,
129 .ptr___pthread_rwlock_wrlock = __pthread_rwlock_wrlock_internal,
130 .ptr___pthread_rwlock_unlock = __pthread_rwlock_unlock_internal,
131 .ptr___pthread_key_create = __pthread_key_create_internal,
132 .ptr___pthread_getspecific = __pthread_getspecific_internal,
133 .ptr___pthread_setspecific = __pthread_setspecific_internal,
134 .ptr__pthread_cleanup_push_defer = __pthread_cleanup_push_defer,
135 .ptr__pthread_cleanup_pop_restore = __pthread_cleanup_pop_restore,
136 .ptr_nthreads = &__nptl_nthreads,
137 .ptr___pthread_unwind = &__pthread_unwind,
138 .ptr__nptl_deallocate_tsd = __nptl_deallocate_tsd,
139 .ptr__nptl_setxid = __nptl_setxid,
140 /* For now only the stack cache needs to be freed. */
141 .ptr_freeres = nptl_freeres,
142 .ptr_set_robust = __nptl_set_robust
144 # define ptr_pthread_functions &pthread_functions
145 #else
146 # define ptr_pthread_functions NULL
147 #endif
150 #ifdef SHARED
151 /* This function is called indirectly from the freeres code in libc. */
152 static void
153 __libc_freeres_fn_section
154 nptl_freeres (void)
156 __unwind_freeres ();
157 __free_stacks (0);
161 static
162 #endif
163 void
164 __nptl_set_robust (struct pthread *self)
166 INTERNAL_SYSCALL_DECL (err);
167 INTERNAL_SYSCALL (set_robust_list, err, 2, &self->robust_head,
168 sizeof (struct robust_list_head));
172 /* For asynchronous cancellation we use a signal. This is the handler. */
173 static void
174 sigcancel_handler (int sig, siginfo_t *si, void *ctx)
176 #ifdef __ASSUME_CORRECT_SI_PID
177 /* Determine the process ID. It might be negative if the thread is
178 in the middle of a fork() call. */
179 pid_t pid = THREAD_GETMEM (THREAD_SELF, pid);
180 if (__builtin_expect (pid < 0, 0))
181 pid = -pid;
182 #endif
184 /* Safety check. It would be possible to call this function for
185 other signals and send a signal from another process. This is not
186 correct and might even be a security problem. Try to catch as
187 many incorrect invocations as possible. */
188 if (sig != SIGCANCEL
189 #ifdef __ASSUME_CORRECT_SI_PID
190 /* Kernels before 2.5.75 stored the thread ID and not the process
191 ID in si_pid so we skip this test. */
192 || si->si_pid != pid
193 #endif
194 || si->si_code != SI_TKILL)
195 return;
197 struct pthread *self = THREAD_SELF;
199 int oldval = THREAD_GETMEM (self, cancelhandling);
200 while (1)
202 /* We are canceled now. When canceled by another thread this flag
203 is already set but if the signal is directly send (internally or
204 from another process) is has to be done here. */
205 int newval = oldval | CANCELING_BITMASK | CANCELED_BITMASK;
207 if (oldval == newval || (oldval & EXITING_BITMASK) != 0)
208 /* Already canceled or exiting. */
209 break;
211 int curval = THREAD_ATOMIC_CMPXCHG_VAL (self, cancelhandling, newval,
212 oldval);
213 if (curval == oldval)
215 /* Set the return value. */
216 THREAD_SETMEM (self, result, PTHREAD_CANCELED);
218 /* Make sure asynchronous cancellation is still enabled. */
219 if ((newval & CANCELTYPE_BITMASK) != 0)
220 /* Run the registered destructors and terminate the thread. */
221 __do_cancel ();
223 break;
226 oldval = curval;
231 struct xid_command *__xidcmd attribute_hidden;
233 /* For asynchronous cancellation we use a signal. This is the handler. */
234 static void
235 sighandler_setxid (int sig, siginfo_t *si, void *ctx)
237 #ifdef __ASSUME_CORRECT_SI_PID
238 /* Determine the process ID. It might be negative if the thread is
239 in the middle of a fork() call. */
240 pid_t pid = THREAD_GETMEM (THREAD_SELF, pid);
241 if (__builtin_expect (pid < 0, 0))
242 pid = -pid;
243 #endif
245 /* Safety check. It would be possible to call this function for
246 other signals and send a signal from another process. This is not
247 correct and might even be a security problem. Try to catch as
248 many incorrect invocations as possible. */
249 if (sig != SIGSETXID
250 #ifdef __ASSUME_CORRECT_SI_PID
251 /* Kernels before 2.5.75 stored the thread ID and not the process
252 ID in si_pid so we skip this test. */
253 || si->si_pid != pid
254 #endif
255 || si->si_code != SI_TKILL)
256 return;
258 INTERNAL_SYSCALL_DECL (err);
259 INTERNAL_SYSCALL_NCS (__xidcmd->syscall_no, err, 3, __xidcmd->id[0],
260 __xidcmd->id[1], __xidcmd->id[2]);
262 /* Reset the SETXID flag. */
263 struct pthread *self = THREAD_SELF;
264 int flags, newval;
267 flags = THREAD_GETMEM (self, cancelhandling);
268 newval = THREAD_ATOMIC_CMPXCHG_VAL (self, cancelhandling,
269 flags & ~SETXID_BITMASK, flags);
271 while (flags != newval);
273 /* And release the futex. */
274 self->setxid_futex = 1;
275 lll_futex_wake (&self->setxid_futex, 1, LLL_PRIVATE);
277 if (atomic_decrement_val (&__xidcmd->cntr) == 0)
278 lll_futex_wake (&__xidcmd->cntr, 1, LLL_PRIVATE);
282 /* When using __thread for this, we do it in libc so as not
283 to give libpthread its own TLS segment just for this. */
284 extern void **__libc_dl_error_tsd (void) __attribute__ ((const));
287 /* This can be set by the debugger before initialization is complete. */
288 static bool __nptl_initial_report_events __attribute_used__;
290 void
291 __pthread_initialize_minimal_internal (void)
293 #ifndef SHARED
294 /* Unlike in the dynamically linked case the dynamic linker has not
295 taken care of initializing the TLS data structures. */
296 __libc_setup_tls (TLS_TCB_SIZE, TLS_TCB_ALIGN);
298 /* We must prevent gcc from being clever and move any of the
299 following code ahead of the __libc_setup_tls call. This function
300 will initialize the thread register which is subsequently
301 used. */
302 __asm __volatile ("");
303 #endif
305 /* Minimal initialization of the thread descriptor. */
306 struct pthread *pd = THREAD_SELF;
307 INTERNAL_SYSCALL_DECL (err);
308 pd->pid = pd->tid = INTERNAL_SYSCALL (set_tid_address, err, 1, &pd->tid);
309 THREAD_SETMEM (pd, specific[0], &pd->specific_1stblock[0]);
310 THREAD_SETMEM (pd, user_stack, true);
311 if (LLL_LOCK_INITIALIZER != 0)
312 THREAD_SETMEM (pd, lock, LLL_LOCK_INITIALIZER);
313 #if HP_TIMING_AVAIL
314 THREAD_SETMEM (pd, cpuclock_offset, GL(dl_cpuclock_offset));
315 #endif
317 /* Initialize the robust mutex data. */
318 #ifdef __PTHREAD_MUTEX_HAVE_PREV
319 pd->robust_prev = &pd->robust_head;
320 #endif
321 pd->robust_head.list = &pd->robust_head;
322 #ifdef __NR_set_robust_list
323 pd->robust_head.futex_offset = (offsetof (pthread_mutex_t, __data.__lock)
324 - offsetof (pthread_mutex_t,
325 __data.__list.__next));
326 int res = INTERNAL_SYSCALL (set_robust_list, err, 2, &pd->robust_head,
327 sizeof (struct robust_list_head));
328 if (INTERNAL_SYSCALL_ERROR_P (res, err))
329 #endif
330 set_robust_list_not_avail ();
332 #ifndef __ASSUME_PRIVATE_FUTEX
333 /* Private futexes are always used (at least internally) so that
334 doing the test once this early is beneficial. */
336 int word = 0;
337 word = INTERNAL_SYSCALL (futex, err, 3, &word,
338 FUTEX_WAKE | FUTEX_PRIVATE_FLAG, 1);
339 if (!INTERNAL_SYSCALL_ERROR_P (word, err))
340 THREAD_SETMEM (pd, header.private_futex, FUTEX_PRIVATE_FLAG);
343 /* Private futexes have been introduced earlier than the
344 FUTEX_CLOCK_REALTIME flag. We don't have to run the test if we
345 know the former are not supported. This also means we know the
346 kernel will return ENOSYS for unknown operations. */
347 if (THREAD_GETMEM (pd, header.private_futex) != 0)
348 #endif
349 #ifndef __ASSUME_FUTEX_CLOCK_REALTIME
351 int word = 0;
352 /* NB: the syscall actually takes six parameters. The last is the
353 bit mask. But since we will not actually wait at all the value
354 is irrelevant. Given that passing six parameters is difficult
355 on some architectures we just pass whatever random value the
356 calling convention calls for to the kernel. It causes no harm. */
357 word = INTERNAL_SYSCALL (futex, err, 5, &word,
358 FUTEX_WAIT_BITSET | FUTEX_CLOCK_REALTIME
359 | FUTEX_PRIVATE_FLAG, 1, NULL, 0);
360 assert (INTERNAL_SYSCALL_ERROR_P (word, err));
361 if (INTERNAL_SYSCALL_ERRNO (word, err) != ENOSYS)
362 __set_futex_clock_realtime ();
364 #endif
366 /* Set initial thread's stack block from 0 up to __libc_stack_end.
367 It will be bigger than it actually is, but for unwind.c/pt-longjmp.c
368 purposes this is good enough. */
369 THREAD_SETMEM (pd, stackblock_size, (size_t) __libc_stack_end);
371 /* Initialize the list of all running threads with the main thread. */
372 INIT_LIST_HEAD (&__stack_user);
373 list_add (&pd->list, &__stack_user);
375 /* Before initializing __stack_user, the debugger could not find us and
376 had to set __nptl_initial_report_events. Propagate its setting. */
377 THREAD_SETMEM (pd, report_events, __nptl_initial_report_events);
379 /* Install the cancellation signal handler. If for some reason we
380 cannot install the handler we do not abort. Maybe we should, but
381 it is only asynchronous cancellation which is affected. */
382 struct sigaction sa;
383 sa.sa_sigaction = sigcancel_handler;
384 sa.sa_flags = SA_SIGINFO;
385 __sigemptyset (&sa.sa_mask);
387 (void) __libc_sigaction (SIGCANCEL, &sa, NULL);
389 /* Install the handle to change the threads' uid/gid. */
390 sa.sa_sigaction = sighandler_setxid;
391 sa.sa_flags = SA_SIGINFO | SA_RESTART;
393 (void) __libc_sigaction (SIGSETXID, &sa, NULL);
395 /* The parent process might have left the signals blocked. Just in
396 case, unblock it. We reuse the signal mask in the sigaction
397 structure. It is already cleared. */
398 __sigaddset (&sa.sa_mask, SIGCANCEL);
399 __sigaddset (&sa.sa_mask, SIGSETXID);
400 (void) INTERNAL_SYSCALL (rt_sigprocmask, err, 4, SIG_UNBLOCK, &sa.sa_mask,
401 NULL, _NSIG / 8);
403 /* Get the size of the static and alignment requirements for the TLS
404 block. */
405 size_t static_tls_align;
406 _dl_get_tls_static_info (&__static_tls_size, &static_tls_align);
408 /* Make sure the size takes all the alignments into account. */
409 if (STACK_ALIGN > static_tls_align)
410 static_tls_align = STACK_ALIGN;
411 __static_tls_align_m1 = static_tls_align - 1;
413 __static_tls_size = roundup (__static_tls_size, static_tls_align);
415 /* Determine the default allowed stack size. This is the size used
416 in case the user does not specify one. */
417 struct rlimit limit;
418 if (getrlimit (RLIMIT_STACK, &limit) != 0
419 || limit.rlim_cur == RLIM_INFINITY)
420 /* The system limit is not usable. Use an architecture-specific
421 default. */
422 limit.rlim_cur = ARCH_STACK_DEFAULT_SIZE;
423 else if (limit.rlim_cur < PTHREAD_STACK_MIN)
424 /* The system limit is unusably small.
425 Use the minimal size acceptable. */
426 limit.rlim_cur = PTHREAD_STACK_MIN;
428 /* Make sure it meets the minimum size that allocate_stack
429 (allocatestack.c) will demand, which depends on the page size. */
430 const uintptr_t pagesz = __sysconf (_SC_PAGESIZE);
431 const size_t minstack = pagesz + __static_tls_size + MINIMAL_REST_STACK;
432 if (limit.rlim_cur < minstack)
433 limit.rlim_cur = minstack;
435 /* Round the resource limit up to page size. */
436 limit.rlim_cur = (limit.rlim_cur + pagesz - 1) & -pagesz;
437 __default_stacksize = limit.rlim_cur;
439 #ifdef SHARED
440 /* Transfer the old value from the dynamic linker's internal location. */
441 *__libc_dl_error_tsd () = *(*GL(dl_error_catch_tsd)) ();
442 GL(dl_error_catch_tsd) = &__libc_dl_error_tsd;
444 /* Make __rtld_lock_{,un}lock_recursive use pthread_mutex_{,un}lock,
445 keep the lock count from the ld.so implementation. */
446 GL(dl_rtld_lock_recursive) = (void *) INTUSE (__pthread_mutex_lock);
447 GL(dl_rtld_unlock_recursive) = (void *) INTUSE (__pthread_mutex_unlock);
448 unsigned int rtld_lock_count = GL(dl_load_lock).mutex.__data.__count;
449 GL(dl_load_lock).mutex.__data.__count = 0;
450 while (rtld_lock_count-- > 0)
451 INTUSE (__pthread_mutex_lock) (&GL(dl_load_lock).mutex);
453 GL(dl_make_stack_executable_hook) = &__make_stacks_executable;
454 #endif
456 GL(dl_init_static_tls) = &__pthread_init_static_tls;
458 GL(dl_wait_lookup_done) = &__wait_lookup_done;
460 /* Register the fork generation counter with the libc. */
461 #ifndef TLS_MULTIPLE_THREADS_IN_TCB
462 __libc_multiple_threads_ptr =
463 #endif
464 __libc_pthread_init (&__fork_generation, __reclaim_stacks,
465 ptr_pthread_functions);
467 /* Determine whether the machine is SMP or not. */
468 __is_smp = is_smp_system ();
470 strong_alias (__pthread_initialize_minimal_internal,
471 __pthread_initialize_minimal)