+#define _GNU_SOURCE
#include "pthread_impl.h"
+#include "stdio_impl.h"
+#include "libc.h"
+#include <sys/mman.h>
+#include <string.h>
+#include <stddef.h>
-static void dummy_1(pthread_t self)
+void *__mmap(void *, size_t, int, int, int, off_t);
+int __munmap(void *, size_t);
+int __mprotect(void *, size_t, int);
+
+static void dummy_0()
{
}
-weak_alias(dummy_1, __pthread_tsd_run_dtors);
-weak_alias(dummy_1, __sigtimer_handler);
-
-#ifdef __pthread_unwind_next
-#undef __pthread_unwind_next
-#define __pthread_unwind_next __pthread_unwind_next_3
-#endif
+weak_alias(dummy_0, __acquire_ptc);
+weak_alias(dummy_0, __release_ptc);
+weak_alias(dummy_0, __pthread_tsd_run_dtors);
+weak_alias(dummy_0, __do_orphaned_stdio_locks);
+weak_alias(dummy_0, __dl_thread_cleanup);
-void __pthread_unwind_next(struct __ptcb *cb)
+_Noreturn void __pthread_exit(void *result)
{
- pthread_t self;
+ pthread_t self = __pthread_self();
+ sigset_t set;
- if (cb->__next) longjmp((void *)cb->__next->__jb, 1);
+ self->canceldisable = 1;
+ self->cancelasync = 0;
+ self->result = result;
- self = pthread_self();
+ while (self->cancelbuf) {
+ void (*f)(void *) = self->cancelbuf->__f;
+ void *x = self->cancelbuf->__x;
+ self->cancelbuf = self->cancelbuf->__next;
+ f(x);
+ }
- LOCK(&self->exitlock);
+ __pthread_tsd_run_dtors();
- __pthread_tsd_run_dtors(self);
+ __lock(self->exitlock);
/* Mark this thread dead before decrementing count */
+ __lock(self->killlock);
self->dead = 1;
- if (!a_fetch_add(&libc.threads_minus_1, -1))
+ /* Block all signals before decrementing the live thread count.
+ * This is important to ensure that dynamically allocated TLS
+ * is not under-allocated/over-committed, and possibly for other
+ * reasons as well. */
+ __block_all_sigs(&set);
+
+ /* Wait to unlock the kill lock, which governs functions like
+ * pthread_kill which target a thread id, until signals have
+ * been blocked. This precludes observation of the thread id
+ * as a live thread (with application code running in it) after
+ * the thread was reported dead by ESRCH being returned. */
+ __unlock(self->killlock);
+
+ /* It's impossible to determine whether this is "the last thread"
+ * until performing the atomic decrement, since multiple threads
+ * could exit at the same time. For the last thread, revert the
+ * decrement and unblock signals to give the atexit handlers and
+ * stdio cleanup code a consistent state. */
+ if (a_fetch_add(&libc.threads_minus_1, -1)==0) {
+ libc.threads_minus_1 = 0;
+ __restore_sigs(&set);
exit(0);
+ }
+
+ /* Process robust list in userspace to handle non-pshared mutexes
+ * and the detached thread case where the robust list head will
+ * be invalid when the kernel would process it. */
+ __vm_lock();
+ volatile void *volatile *rp;
+ while ((rp=self->robust_list.head) && rp != &self->robust_list.head) {
+ pthread_mutex_t *m = (void *)((char *)rp
+ - offsetof(pthread_mutex_t, _m_next));
+ int waiters = m->_m_waiters;
+ int priv = (m->_m_type & 128) ^ 128;
+ self->robust_list.pending = rp;
+ self->robust_list.head = *rp;
+ int cont = a_swap(&m->_m_lock, self->tid|0x40000000);
+ self->robust_list.pending = 0;
+ if (cont < 0 || waiters)
+ __wake(&m->_m_lock, 1, priv);
+ }
+ __vm_unlock();
+
+ __do_orphaned_stdio_locks();
+ __dl_thread_cleanup();
if (self->detached && self->map_base) {
- syscall(__NR_rt_sigprocmask, SIG_BLOCK, (long)(uint64_t[1]){-1},0,8);
+ /* Detached threads must avoid the kernel clear_child_tid
+ * feature, since the virtual address will have been
+ * unmapped and possibly already reused by a new mapping
+ * at the time the kernel would perform the write. In
+ * the case of threads that started out detached, the
+ * initial clone flags are correct, but if the thread was
+ * detached later (== 2), we need to clear it here. */
+ if (self->detached == 2) __syscall(SYS_set_tid_address, 0);
+
+ /* Robust list will no longer be valid, and was already
+ * processed above, so unregister it with the kernel. */
+ if (self->robust_list.off)
+ __syscall(SYS_set_robust_list, 0, 3*sizeof(long));
+
+ /* Since __unmapself bypasses the normal munmap code path,
+ * explicitly wait for vmlock holders first. */
+ __vm_wait();
+
+ /* The following call unmaps the thread's stack mapping
+ * and then exits without touching the stack. */
__unmapself(self->map_base, self->map_size);
}
- syscall(SYS_exit, 0);
-}
-
-static void docancel(struct pthread *self)
-{
- struct __ptcb cb = { .__next = self->cancelbuf };
- self->canceldisable = 1;
- self->cancelasync = 0;
- __pthread_unwind_next(&cb);
+ for (;;) __syscall(SYS_exit, 0);
}
-static void cancel_handler(int sig, siginfo_t *si, void *ctx)
+void __do_cleanup_push(struct __ptcb *cb)
{
struct pthread *self = __pthread_self();
- if (si->si_code == SI_TIMER) __sigtimer_handler(self);
- if (self->cancel && !self->canceldisable &&
- (self->cancelasync || (self->cancelpoint==1 && PC_AT_SYS(ctx))))
- docancel(self);
+ cb->__next = self->cancelbuf;
+ self->cancelbuf = cb;
}
-static void cancelpt(int x)
+void __do_cleanup_pop(struct __ptcb *cb)
{
- struct pthread *self = __pthread_self();
- if (self->canceldisable) return;
- if ((self->cancelpoint+=x)==1 && x>=0 && self->cancel)
- docancel(self);
+ __pthread_self()->cancelbuf = cb->__next;
}
-/* "rsyscall" is a mechanism by which a thread can synchronously force all
- * other threads to perform an arbitrary syscall. It is necessary to work
- * around the non-conformant implementation of setuid() et al on Linux,
- * which affect only the calling thread and not the whole process. This
- * implementation performs some tricks with signal delivery to work around
- * the fact that it does not keep any list of threads in userspace. */
-
-static struct {
- volatile int lock, hold, blocks, cnt;
- unsigned long arg[6];
- int nr;
- int err;
-} rs;
-
-static void rsyscall_handler(int sig, siginfo_t *si, void *ctx)
+static int start(void *p)
{
- struct pthread *self = __pthread_self();
-
- if (!rs.hold || rs.cnt == libc.threads_minus_1) return;
-
- /* Threads which have already decremented themselves from the
- * thread count must not increment rs.cnt or otherwise act. */
- if (self->dead) {
- sigfillset(&((ucontext_t *)ctx)->uc_sigmask);
- return;
+ pthread_t self = p;
+ if (self->startlock[0]) {
+ __wait(self->startlock, 0, 1, 1);
+ if (self->startlock[0]) {
+ self->detached = 2;
+ pthread_exit(0);
+ }
+ __restore_sigs(self->sigmask);
}
-
- if (syscall(rs.nr, rs.arg[0], rs.arg[1], rs.arg[2],
- rs.arg[3], rs.arg[4], rs.arg[5]) < 0 && !rs.err) rs.err=errno;
-
- a_inc(&rs.cnt);
- __wake(&rs.cnt, 1, 1);
- while(rs.hold)
- __wait(&rs.hold, 0, 1, 1);
- a_dec(&rs.cnt);
- if (!rs.cnt) __wake(&rs.cnt, 1, 1);
+ if (self->unblock_cancel)
+ __syscall(SYS_rt_sigprocmask, SIG_UNBLOCK,
+ SIGPT_SET, 0, _NSIG/8);
+ __pthread_exit(self->start(self->start_arg));
+ return 0;
}
-static int rsyscall(int nr, long a, long b, long c, long d, long e, long f)
+static int start_c11(void *p)
{
- int i, ret;
- sigset_t set = { 0 };
- struct pthread *self = __pthread_self();
- sigaddset(&set, SIGSYSCALL);
-
- LOCK(&rs.lock);
- while ((i=rs.blocks))
- __wait(&rs.blocks, 0, i, 1);
-
- __libc_sigprocmask(SIG_BLOCK, &set, 0);
-
- rs.nr = nr;
- rs.arg[0] = a; rs.arg[1] = b;
- rs.arg[2] = c; rs.arg[3] = d;
- rs.arg[4] = d; rs.arg[5] = f;
- rs.err = 0;
- rs.cnt = 0;
- rs.hold = 1;
-
- /* Dispatch signals until all threads respond */
- for (i=libc.threads_minus_1; i; i--)
- sigqueue(self->pid, SIGSYSCALL, (union sigval){0});
- while ((i=rs.cnt) < libc.threads_minus_1) {
- sigqueue(self->pid, SIGSYSCALL, (union sigval){0});
- __wait(&rs.cnt, 0, i, 1);
- }
+ pthread_t self = p;
+ int (*start)(void*) = (int(*)(void*)) self->start;
+ __pthread_exit((void *)(uintptr_t)start(self->start_arg));
+ return 0;
+}
- /* Handle any lingering signals with no-op */
- __libc_sigprocmask(SIG_UNBLOCK, &set, 0);
+#define ROUND(x) (((x)+PAGE_SIZE-1)&-PAGE_SIZE)
- /* Resume other threads' signal handlers and wait for them */
- rs.hold = 0;
- __wake(&rs.hold, -1, 0);
- while((i=rs.cnt)) __wait(&rs.cnt, 0, i, 1);
+/* pthread_key_create.c overrides this */
+static volatile size_t dummy = 0;
+weak_alias(dummy, __pthread_tsd_size);
+static void *dummy_tsd[1] = { 0 };
+weak_alias(dummy_tsd, __pthread_tsd_main);
- if (rs.err) errno = rs.err, ret = -1;
- else ret = syscall(nr, a, b, c, d, e, f);
+volatile int __block_new_threads = 0;
- UNLOCK(&rs.lock);
- return ret;
-}
+static FILE *volatile dummy_file = 0;
+weak_alias(dummy_file, __stdin_used);
+weak_alias(dummy_file, __stdout_used);
+weak_alias(dummy_file, __stderr_used);
-static void init_threads()
+static void init_file_lock(FILE *f)
{
- struct sigaction sa = { .sa_flags = SA_SIGINFO | SA_RESTART };
- libc.lock = __lock;
- libc.lockfile = __lockfile;
- libc.cancelpt = cancelpt;
- libc.rsyscall = rsyscall;
-
- sigfillset(&sa.sa_mask);
- sa.sa_sigaction = rsyscall_handler;
- __libc_sigaction(SIGSYSCALL, &sa, 0);
-
- sigemptyset(&sa.sa_mask);
- sa.sa_sigaction = cancel_handler;
- __libc_sigaction(SIGCANCEL, &sa, 0);
-
- sigaddset(&sa.sa_mask, SIGSYSCALL);
- sigaddset(&sa.sa_mask, SIGCANCEL);
- __libc_sigprocmask(SIG_UNBLOCK, &sa.sa_mask, 0);
+ if (f && f->lock<0) f->lock = 0;
}
-static int start(void *p)
+void *__copy_tls(unsigned char *);
+
+int __pthread_create(pthread_t *restrict res, const pthread_attr_t *restrict attrp, void *(*entry)(void *), void *restrict arg)
{
- struct pthread *self = p;
- if (self->unblock_cancel) {
- sigset_t set;
- sigemptyset(&set);
- sigaddset(&set, SIGCANCEL);
- __libc_sigprocmask(SIG_UNBLOCK, &set, 0);
+ int ret, c11 = (attrp == __ATTRP_C11_THREAD);
+ size_t size, guard;
+ struct pthread *self, *new;
+ unsigned char *map = 0, *stack = 0, *tsd = 0, *stack_limit;
+ unsigned flags = CLONE_VM | CLONE_FS | CLONE_FILES | CLONE_SIGHAND
+ | CLONE_THREAD | CLONE_SYSVSEM | CLONE_SETTLS
+ | CLONE_PARENT_SETTID | CLONE_CHILD_CLEARTID | CLONE_DETACHED;
+ int do_sched = 0;
+ pthread_attr_t attr = {0};
+
+ if (!libc.can_do_threads) return ENOSYS;
+ self = __pthread_self();
+ if (!libc.threaded) {
+ for (FILE *f=*__ofl_lock(); f; f=f->next)
+ init_file_lock(f);
+ __ofl_unlock();
+ init_file_lock(__stdin_used);
+ init_file_lock(__stdout_used);
+ init_file_lock(__stderr_used);
+ __syscall(SYS_rt_sigprocmask, SIG_UNBLOCK, SIGPT_SET, 0, _NSIG/8);
+ self->tsd = (void **)__pthread_tsd_main;
+ libc.threaded = 1;
+ }
+ if (attrp && !c11) attr = *attrp;
+
+ __acquire_ptc();
+ if (__block_new_threads) __wait(&__block_new_threads, 0, 1, 1);
+
+ if (attr._a_stackaddr) {
+ size_t need = libc.tls_size + __pthread_tsd_size;
+ size = attr._a_stacksize + DEFAULT_STACK_SIZE;
+ stack = (void *)(attr._a_stackaddr & -16);
+ stack_limit = (void *)(attr._a_stackaddr - size);
+ /* Use application-provided stack for TLS only when
+ * it does not take more than ~12% or 2k of the
+ * application's stack space. */
+ if (need < size/8 && need < 2048) {
+ tsd = stack - __pthread_tsd_size;
+ stack = tsd - libc.tls_size;
+ memset(stack, 0, need);
+ } else {
+ size = ROUND(need);
+ guard = 0;
+ }
+ } else {
+ guard = ROUND(DEFAULT_GUARD_SIZE + attr._a_guardsize);
+ size = guard + ROUND(DEFAULT_STACK_SIZE + attr._a_stacksize
+ + libc.tls_size + __pthread_tsd_size);
}
- pthread_exit(self->start(self->start_arg));
- return 0;
-}
-
-int __uniclone(void *, int (*)(), void *);
-
-#define ROUND(x) (((x)+PAGE_SIZE-1)&-PAGE_SIZE)
-/* pthread_key_create.c overrides this */
-static const size_t dummy = 0;
-weak_alias(dummy, __pthread_tsd_size);
+ if (!tsd) {
+ if (guard) {
+ map = __mmap(0, size, PROT_NONE, MAP_PRIVATE|MAP_ANON, -1, 0);
+ if (map == MAP_FAILED) goto fail;
+ if (__mprotect(map+guard, size-guard, PROT_READ|PROT_WRITE)) {
+ __munmap(map, size);
+ goto fail;
+ }
+ } else {
+ map = __mmap(0, size, PROT_READ|PROT_WRITE, MAP_PRIVATE|MAP_ANON, -1, 0);
+ if (map == MAP_FAILED) goto fail;
+ }
+ tsd = map + size - __pthread_tsd_size;
+ if (!stack) {
+ stack = tsd - libc.tls_size;
+ stack_limit = map + guard;
+ }
+ }
-int pthread_create(pthread_t *res, const pthread_attr_t *attr, void *(*entry)(void *), void *arg)
-{
- static int init;
- int ret;
- size_t size, guard;
- struct pthread *self = pthread_self(), *new;
- unsigned char *map, *stack, *tsd;
- static const pthread_attr_t default_attr;
-
- if (!self) return errno = ENOSYS;
- if (!init && ++init) init_threads();
-
- if (!attr) attr = &default_attr;
- guard = ROUND(attr->_a_guardsize + DEFAULT_GUARD_SIZE);
- size = guard + ROUND(attr->_a_stacksize + DEFAULT_STACK_SIZE);
- size += __pthread_tsd_size;
- map = mmap(0, size, PROT_READ|PROT_WRITE|PROT_EXEC, MAP_PRIVATE|MAP_ANON, -1, 0);
- if (!map) return EAGAIN;
- if (guard) mprotect(map, guard, PROT_NONE);
-
- tsd = map + size - __pthread_tsd_size;
- new = (void *)(tsd - sizeof *new - PAGE_SIZE%sizeof *new);
+ new = __copy_tls(tsd - libc.tls_size);
new->map_base = map;
new->map_size = size;
- new->pid = self->pid;
- new->errno_ptr = &new->errno_val;
+ new->stack = stack;
+ new->stack_size = stack - stack_limit;
new->start = entry;
new->start_arg = arg;
new->self = new;
new->tsd = (void *)tsd;
- new->detached = attr->_a_detach;
- new->attr = *attr;
+ new->locale = &libc.global_locale;
+ if (attr._a_detach) {
+ new->detached = 1;
+ flags -= CLONE_CHILD_CLEARTID;
+ }
+ if (attr._a_sched) {
+ do_sched = new->startlock[0] = 1;
+ __block_app_sigs(new->sigmask);
+ }
+ new->robust_list.head = &new->robust_list.head;
new->unblock_cancel = self->cancel;
- new->result = PTHREAD_CANCELED;
- memcpy(new->tlsdesc, self->tlsdesc, sizeof new->tlsdesc);
- new->tlsdesc[1] = (uintptr_t)new;
- stack = (void *)((uintptr_t)new-1 & ~(uintptr_t)15);
-
- /* We must synchronize new thread creation with rsyscall
- * delivery. This looks to be the least expensive way: */
- a_inc(&rs.blocks);
- while (rs.lock) __wait(&rs.lock, 0, 1, 1);
+ new->CANARY = self->CANARY;
a_inc(&libc.threads_minus_1);
- ret = __uniclone(stack, start, new);
+ ret = __clone((c11 ? start_c11 : start), stack, flags, new, &new->tid, TP_ADJ(new), &new->tid);
+
+ __release_ptc();
- a_dec(&rs.blocks);
- if (rs.lock) __wake(&rs.blocks, 1, 1);
+ if (do_sched) {
+ __restore_sigs(new->sigmask);
+ }
if (ret < 0) {
a_dec(&libc.threads_minus_1);
- munmap(map, size);
+ if (map) __munmap(map, size);
return EAGAIN;
}
+
+ if (do_sched) {
+ ret = __syscall(SYS_sched_setscheduler, new->tid,
+ attr._a_policy, &attr._a_prio);
+ a_store(new->startlock, ret<0 ? 2 : 0);
+ __wake(new->startlock, 1, 1);
+ if (ret < 0) return -ret;
+ }
+
*res = new;
return 0;
+fail:
+ __release_ptc();
+ return EAGAIN;
}
-void pthread_exit(void *result)
-{
- struct pthread *self = pthread_self();
- self->result = result;
- docancel(self);
-}
+weak_alias(__pthread_exit, pthread_exit);
+weak_alias(__pthread_create, pthread_create);