1#define _GNU_SOURCE
2#include "pthread_impl.h"
3#include "stdio_impl.h"
4#include "libc.h"
5#include <sys/mman.h>
6#include <string.h>
7#include <stddef.h>
8
9void *__mmap(void *, size_t, int, int, int, off_t);
10int __munmap(void *, size_t);
11int __mprotect(void *, size_t, int);
12
13static void dummy_0()
14{
15}
16weak_alias(dummy_0, __acquire_ptc);
17weak_alias(dummy_0, __release_ptc);
18weak_alias(dummy_0, __pthread_tsd_run_dtors);
19weak_alias(dummy_0, __do_orphaned_stdio_locks);
20weak_alias(dummy_0, __dl_thread_cleanup);
21
22_Noreturn void __pthread_exit(void *result)
23{
24	pthread_t self = __pthread_self();
25	sigset_t set;
26
27	self->canceldisable = 1;
28	self->cancelasync = 0;
29	self->result = result;
30
31	while (self->cancelbuf) {
32		void (*f)(void *) = self->cancelbuf->__f;
33		void *x = self->cancelbuf->__x;
34		self->cancelbuf = self->cancelbuf->__next;
35		f(x);
36	}
37
38	__pthread_tsd_run_dtors();
39
40	__lock(self->exitlock);
41
42	/* Mark this thread dead before decrementing count */
43	__lock(self->killlock);
44	self->dead = 1;
45
46	/* Block all signals before decrementing the live thread count.
47	 * This is important to ensure that dynamically allocated TLS
48	 * is not under-allocated/over-committed, and possibly for other
49	 * reasons as well. */
50	__block_all_sigs(&set);
51
52	/* Wait to unlock the kill lock, which governs functions like
53	 * pthread_kill which target a thread id, until signals have
54	 * been blocked. This precludes observation of the thread id
55	 * as a live thread (with application code running in it) after
56	 * the thread was reported dead by ESRCH being returned. */
57	__unlock(self->killlock);
58
59	/* It's impossible to determine whether this is "the last thread"
60	 * until performing the atomic decrement, since multiple threads
61	 * could exit at the same time. For the last thread, revert the
62	 * decrement and unblock signals to give the atexit handlers and
63	 * stdio cleanup code a consistent state. */
64	if (a_fetch_add(&libc.threads_minus_1, -1)==0) {
65		libc.threads_minus_1 = 0;
66		__restore_sigs(&set);
67		exit(0);
68	}
69
70	/* Process robust list in userspace to handle non-pshared mutexes
71	 * and the detached thread case where the robust list head will
72	 * be invalid when the kernel would process it. */
73	__vm_lock();
74	volatile void *volatile *rp;
75	while ((rp=self->robust_list.head) && rp != &self->robust_list.head) {
76		pthread_mutex_t *m = (void *)((char *)rp
77			- offsetof(pthread_mutex_t, _m_next));
78		int waiters = m->_m_waiters;
79		int priv = (m->_m_type & 128) ^ 128;
80		self->robust_list.pending = rp;
81		self->robust_list.head = *rp;
82		int cont = a_swap(&m->_m_lock, 0x40000000);
83		self->robust_list.pending = 0;
84		if (cont < 0 || waiters)
85			__wake(&m->_m_lock, 1, priv);
86	}
87	__vm_unlock();
88
89	__do_orphaned_stdio_locks();
90	__dl_thread_cleanup();
91
92	if (self->detached && self->map_base) {
93		/* Detached threads must avoid the kernel clear_child_tid
94		 * feature, since the virtual address will have been
95		 * unmapped and possibly already reused by a new mapping
96		 * at the time the kernel would perform the write. In
97		 * the case of threads that started out detached, the
98		 * initial clone flags are correct, but if the thread was
99		 * detached later (== 2), we need to clear it here. */
100		if (self->detached == 2) __syscall(SYS_set_tid_address, 0);
101
102		/* Robust list will no longer be valid, and was already
103		 * processed above, so unregister it with the kernel. */
104		if (self->robust_list.off)
105			__syscall(SYS_set_robust_list, 0, 3*sizeof(long));
106
107		/* Since __unmapself bypasses the normal munmap code path,
108		 * explicitly wait for vmlock holders first. */
109		__vm_wait();
110
111		/* The following call unmaps the thread's stack mapping
112		 * and then exits without touching the stack. */
113		__unmapself(self->map_base, self->map_size);
114	}
115
116	for (;;) __syscall(SYS_exit, 0);
117}
118
119void __do_cleanup_push(struct __ptcb *cb)
120{
121	struct pthread *self = __pthread_self();
122	cb->__next = self->cancelbuf;
123	self->cancelbuf = cb;
124}
125
126void __do_cleanup_pop(struct __ptcb *cb)
127{
128	__pthread_self()->cancelbuf = cb->__next;
129}
130
131static int start(void *p)
132{
133	pthread_t self = p;
134	if (self->startlock[0]) {
135		__wait(self->startlock, 0, 1, 1);
136		if (self->startlock[0]) {
137			self->detached = 2;
138			pthread_exit(0);
139		}
140		__restore_sigs(self->sigmask);
141	}
142	if (self->unblock_cancel)
143		__syscall(SYS_rt_sigprocmask, SIG_UNBLOCK,
144			SIGPT_SET, 0, _NSIG/8);
145	__pthread_exit(self->start(self->start_arg));
146	return 0;
147}
148
149static int start_c11(void *p)
150{
151	pthread_t self = p;
152	int (*start)(void*) = (int(*)(void*)) self->start;
153	__pthread_exit((void *)(uintptr_t)start(self->start_arg));
154	return 0;
155}
156
157#define ROUND(x) (((x)+PAGE_SIZE-1)&-PAGE_SIZE)
158
159/* pthread_key_create.c overrides this */
160static volatile size_t dummy = 0;
161weak_alias(dummy, __pthread_tsd_size);
162static void *dummy_tsd[1] = { 0 };
163weak_alias(dummy_tsd, __pthread_tsd_main);
164
165volatile int __block_new_threads = 0;
166size_t __default_stacksize = DEFAULT_STACK_SIZE;
167size_t __default_guardsize = DEFAULT_GUARD_SIZE;
168
169static FILE *volatile dummy_file = 0;
170weak_alias(dummy_file, __stdin_used);
171weak_alias(dummy_file, __stdout_used);
172weak_alias(dummy_file, __stderr_used);
173
174static void init_file_lock(FILE *f)
175{
176	if (f && f->lock<0) f->lock = 0;
177}
178
179void *__copy_tls(unsigned char *);
180
181int __pthread_create(pthread_t *restrict res, const pthread_attr_t *restrict attrp, void *(*entry)(void *), void *restrict arg)
182{
183	int ret, c11 = (attrp == __ATTRP_C11_THREAD);
184	size_t size, guard;
185	struct pthread *self, *new;
186	unsigned char *map = 0, *stack = 0, *tsd = 0, *stack_limit;
187	unsigned flags = CLONE_VM | CLONE_FS | CLONE_FILES | CLONE_SIGHAND
188		| CLONE_THREAD | CLONE_SYSVSEM | CLONE_SETTLS
189		| CLONE_PARENT_SETTID | CLONE_CHILD_CLEARTID | CLONE_DETACHED;
190	int do_sched = 0;
191	pthread_attr_t attr = { 0 };
192
193	if (!libc.can_do_threads) return ENOSYS;
194	self = __pthread_self();
195	if (!libc.threaded) {
196		for (FILE *f=*__ofl_lock(); f; f=f->next)
197			init_file_lock(f);
198		__ofl_unlock();
199		init_file_lock(__stdin_used);
200		init_file_lock(__stdout_used);
201		init_file_lock(__stderr_used);
202		__syscall(SYS_rt_sigprocmask, SIG_UNBLOCK, SIGPT_SET, 0, _NSIG/8);
203		self->tsd = (void **)__pthread_tsd_main;
204		libc.threaded = 1;
205	}
206	if (attrp && !c11) attr = *attrp;
207
208	__acquire_ptc();
209	if (!attrp || c11) {
210		attr._a_stacksize = __default_stacksize;
211		attr._a_guardsize = __default_guardsize;
212	}
213
214	if (__block_new_threads) __wait(&__block_new_threads, 0, 1, 1);
215
216	if (attr._a_stackaddr) {
217		size_t need = libc.tls_size + __pthread_tsd_size;
218		size = attr._a_stacksize;
219		stack = (void *)(attr._a_stackaddr & -16);
220		stack_limit = (void *)(attr._a_stackaddr - size);
221		/* Use application-provided stack for TLS only when
222		 * it does not take more than ~12% or 2k of the
223		 * application's stack space. */
224		if (need < size/8 && need < 2048) {
225			tsd = stack - __pthread_tsd_size;
226			stack = tsd - libc.tls_size;
227			memset(stack, 0, need);
228		} else {
229			size = ROUND(need);
230			guard = 0;
231		}
232	} else {
233		guard = ROUND(attr._a_guardsize);
234		size = guard + ROUND(attr._a_stacksize
235			+ libc.tls_size +  __pthread_tsd_size);
236	}
237
238	if (!tsd) {
239		if (guard) {
240			map = __mmap(0, size, PROT_NONE, MAP_PRIVATE|MAP_ANON, -1, 0);
241			if (map == MAP_FAILED) goto fail;
242			if (__mprotect(map+guard, size-guard, PROT_READ|PROT_WRITE)
243			    && errno != ENOSYS) {
244				__munmap(map, size);
245				goto fail;
246			}
247		} else {
248			map = __mmap(0, size, PROT_READ|PROT_WRITE, MAP_PRIVATE|MAP_ANON, -1, 0);
249			if (map == MAP_FAILED) goto fail;
250		}
251		tsd = map + size - __pthread_tsd_size;
252		if (!stack) {
253			stack = tsd - libc.tls_size;
254			stack_limit = map + guard;
255		}
256	}
257
258	new = __copy_tls(tsd - libc.tls_size);
259	new->map_base = map;
260	new->map_size = size;
261	new->stack = stack;
262	new->stack_size = stack - stack_limit;
263	new->start = entry;
264	new->start_arg = arg;
265	new->self = new;
266	new->tsd = (void *)tsd;
267	new->locale = &libc.global_locale;
268	if (attr._a_detach) {
269		new->detached = 1;
270		flags -= CLONE_CHILD_CLEARTID;
271	}
272	if (attr._a_sched) {
273		do_sched = new->startlock[0] = 1;
274		__block_app_sigs(new->sigmask);
275	}
276	new->robust_list.head = &new->robust_list.head;
277	new->unblock_cancel = self->cancel;
278	new->CANARY = self->CANARY;
279
280	a_inc(&libc.threads_minus_1);
281	ret = __clone((c11 ? start_c11 : start), stack, flags, new, &new->tid, TP_ADJ(new), &new->tid);
282
283	__release_ptc();
284
285	if (do_sched) {
286		__restore_sigs(new->sigmask);
287	}
288
289	if (ret < 0) {
290		a_dec(&libc.threads_minus_1);
291		if (map) __munmap(map, size);
292		return EAGAIN;
293	}
294
295	if (do_sched) {
296		ret = __syscall(SYS_sched_setscheduler, new->tid,
297			attr._a_policy, &attr._a_prio);
298		a_store(new->startlock, ret<0 ? 2 : 0);
299		__wake(new->startlock, 1, 1);
300		if (ret < 0) return -ret;
301	}
302
303	*res = new;
304	return 0;
305fail:
306	__release_ptc();
307	return EAGAIN;
308}
309
310weak_alias(__pthread_exit, pthread_exit);
311weak_alias(__pthread_create, pthread_create);
312