X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=src%2Fthread%2Fpthread_create.c;h=8761381a4d861b6d9919cd0220d31ffccb17bc87;hb=04335d9260c076cf4d9264bd93dd3b06c237a639;hp=48290d35b38afbd9ee3de844c84040be121870e1;hpb=819006a88b9473872fee91135b06f4e23231d97e;p=musl diff --git a/src/thread/pthread_create.c b/src/thread/pthread_create.c index 48290d35..8761381a 100644 --- a/src/thread/pthread_create.c +++ b/src/thread/pthread_create.c @@ -1,18 +1,28 @@ +#define _GNU_SOURCE #include "pthread_impl.h" #include "stdio_impl.h" +#include "libc.h" +#include "lock.h" +#include +#include +#include static void dummy_0() { } -weak_alias(dummy_0, __synccall_lock); -weak_alias(dummy_0, __synccall_unlock); +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_exit(void *result) +_Noreturn void __pthread_exit(void *result) { - pthread_t self = pthread_self(); - int n; + pthread_t self = __pthread_self(); + sigset_t set; + self->canceldisable = 1; + self->cancelasync = 0; self->result = result; while (self->cancelbuf) { @@ -24,28 +34,92 @@ void pthread_exit(void *result) __pthread_tsd_run_dtors(); - __lock(&self->exitlock); + /* Access to target the exiting thread with syscalls that use + * its kernel tid is controlled by killlock. For detached threads, + * any use past this point would have undefined behavior, but for + * joinable threads it's a valid usage that must be handled. */ + LOCK(self->killlock); - /* Mark this thread dead before decrementing count */ - __lock(&self->killlock); - self->dead = 1; - a_store(&self->killlock, 0); + /* 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); - do n = libc.threads_minus_1; - while (n && a_cas(&libc.threads_minus_1, n, n-1)!=n); - if (!n) exit(0); + /* 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, restore the tid, 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; + UNLOCK(self->killlock); + __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, 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(); + + /* This atomic potentially competes with a concurrent pthread_detach + * call; the loser is responsible for freeing thread resources. */ + int state = a_cas(&self->detach_state, DT_JOINABLE, DT_EXITING); - if (self->detached && self->map_base) { - __syscall(SYS_rt_sigprocmask, SIG_BLOCK, (uint64_t[]){-1},0,8); + if (state>=DT_DETACHED && self->map_base) { + /* 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, we need to clear it here. */ + if (state == DT_DYNAMIC) __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); + /* After the kernel thread exits, its tid may be reused. Clear it + * to prevent inadvertent use and inform functions that would use + * it that it's no longer available. */ + self->tid = 0; + UNLOCK(self->killlock); + + for (;;) __syscall(SYS_exit, 0); } void __do_cleanup_push(struct __ptcb *cb) { - struct pthread *self = pthread_self(); + struct pthread *self = __pthread_self(); cb->__next = self->cancelbuf; self->cancelbuf = cb; } @@ -55,22 +129,52 @@ void __do_cleanup_pop(struct __ptcb *cb) __pthread_self()->cancelbuf = cb->__next; } +struct start_args { + void *(*start_func)(void *); + void *start_arg; + pthread_attr_t *attr; + volatile int *perr; + unsigned long sig_mask[_NSIG/8/sizeof(long)]; +}; + static int start(void *p) { - pthread_t self = p; - if (self->unblock_cancel) - __syscall(SYS_rt_sigprocmask, SIG_UNBLOCK, SIGPT_SET, 0, 8); - pthread_exit(self->start(self->start_arg)); + struct start_args *args = p; + if (args->attr) { + pthread_t self = __pthread_self(); + int ret = -__syscall(SYS_sched_setscheduler, self->tid, + args->attr->_a_policy, &args->attr->_a_prio); + if (a_swap(args->perr, ret)==-2) + __wake(args->perr, 1, 1); + if (ret) { + self->detach_state = DT_DYNAMIC; + __pthread_exit(0); + } + } + __syscall(SYS_rt_sigprocmask, SIG_SETMASK, &args->sig_mask, 0, _NSIG/8); + __pthread_exit(args->start_func(args->start_arg)); + return 0; +} + +static int start_c11(void *p) +{ + struct start_args *args = p; + int (*start)(void*) = (int(*)(void*)) args->start_func; + __pthread_exit((void *)(uintptr_t)start(args->start_arg)); return 0; } #define ROUND(x) (((x)+PAGE_SIZE-1)&-PAGE_SIZE) /* pthread_key_create.c overrides this */ -static const size_t dummy = 0; +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); -static FILE *const dummy_file = 0; +volatile int __block_new_threads = 0; + +static FILE *volatile dummy_file = 0; weak_alias(dummy_file, __stdin_used); weak_alias(dummy_file, __stdout_used); weak_alias(dummy_file, __stderr_used); @@ -80,64 +184,150 @@ static void init_file_lock(FILE *f) if (f && f->lock<0) f->lock = 0; } -int pthread_create(pthread_t *res, const pthread_attr_t *attr, void *(*entry)(void *), void *arg) +int __pthread_create(pthread_t *restrict res, const pthread_attr_t *restrict attrp, void *(*entry)(void *), void *restrict arg) { - int ret; - size_t size = DEFAULT_STACK_SIZE + DEFAULT_GUARD_SIZE; - size_t guard = DEFAULT_GUARD_SIZE; - struct pthread *self = pthread_self(), *new; - unsigned char *map, *stack, *tsd; + 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; + pthread_attr_t attr = { 0 }; + sigset_t set; + volatile int err = -1; - if (!self) return ENOSYS; + if (!libc.can_do_threads) return ENOSYS; + self = __pthread_self(); if (!libc.threaded) { - for (FILE *f=libc.ofl_head; f; f=f->next) + 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; - if (attr && attr->_a_stackaddr) { - map = 0; - tsd = (void *)(attr->_a_stackaddr-__pthread_tsd_size & -16); + __acquire_ptc(); + if (!attrp || c11) { + attr._a_stacksize = __default_stacksize; + attr._a_guardsize = __default_guardsize; + } + + 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; + 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 { - if (attr) { - guard = ROUND(attr->_a_guardsize + DEFAULT_GUARD_SIZE); - size = guard + ROUND(attr->_a_stacksize + DEFAULT_STACK_SIZE); + guard = ROUND(attr._a_guardsize); + size = guard + ROUND(attr._a_stacksize + + libc.tls_size + __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) + && errno != ENOSYS) { + __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; } - size += __pthread_tsd_size; - map = mmap(0, size, PROT_READ|PROT_WRITE, MAP_PRIVATE|MAP_ANON, -1, 0); - if (map == MAP_FAILED) return EAGAIN; - if (guard) mprotect(map, guard, PROT_NONE); tsd = map + size - __pthread_tsd_size; + if (!stack) { + stack = tsd - libc.tls_size; + stack_limit = map + guard; + } } - 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->start = entry; - new->start_arg = arg; + new->stack = stack; + new->stack_size = stack - stack_limit; + new->guard_size = guard; new->self = new; new->tsd = (void *)tsd; - if (attr) new->detached = attr->_a_detach; - new->unblock_cancel = self->cancel; - new->canary = self->canary ^ (uintptr_t)&new; - stack = (void *)new; + new->locale = &libc.global_locale; + if (attr._a_detach) { + new->detach_state = DT_DETACHED; + flags -= CLONE_CHILD_CLEARTID; + } else { + new->detach_state = DT_JOINABLE; + } + new->robust_list.head = &new->robust_list.head; + new->CANARY = self->CANARY; - __synccall_lock(); + /* Setup argument structure for the new thread on its stack. */ + stack -= (uintptr_t)stack % sizeof(uintptr_t); + stack -= sizeof(struct start_args); + struct start_args *args = (void *)stack; + args->start_func = entry; + args->start_arg = arg; + if (attr._a_sched) { + args->attr = &attr; + args->perr = &err; + } else { + args->attr = 0; + args->perr = 0; + } + + __block_app_sigs(&set); + + /* Ensure SIGCANCEL is unblocked in new thread. This requires + * working with a copy of the set so we can restore the + * original mask in the calling thread. */ + memcpy(&args->sig_mask, &set, sizeof args->sig_mask); + args->sig_mask[(SIGCANCEL-1)/8/sizeof(long)] &= + ~(1UL<<((SIGCANCEL-1)%(8*sizeof(long)))); a_inc(&libc.threads_minus_1); - ret = __clone(start, stack, 0x7d8f00, new, &new->tid, new, &new->tid); + ret = __clone((c11 ? start_c11 : start), stack, flags, args, &new->tid, TP_ADJ(new), &new->detach_state); - __synccall_unlock(); + __restore_sigs(&set); + __release_ptc(); if (ret < 0) { a_dec(&libc.threads_minus_1); - munmap(map, size); + if (map) __munmap(map, size); return EAGAIN; } + + if (attr._a_sched) { + if (a_cas(&err, -1, -2)==-1) + __wait(&err, 0, -2, 1); + ret = err; + if (ret) return ret; + } + *res = new; return 0; +fail: + __release_ptc(); + return EAGAIN; } + +weak_alias(__pthread_exit, pthread_exit); +weak_alias(__pthread_create, pthread_create);