X-Git-Url: http://nsz.repo.hu/git/?p=musl;a=blobdiff_plain;f=src%2Fthread%2Fpthread_create.c;h=0567f966e02d26fbfc366631af9919e20ce7378c;hp=e67616e7464475187a138017cc9a09d428453721;hb=c3a6839ce95c0f93bd73f7fbe5fcaee4054e5b62;hpb=9ec4283b28cf676292fd5c6f681bef1e90e30c18 diff --git a/src/thread/pthread_create.c b/src/thread/pthread_create.c index e67616e7..0567f966 100644 --- a/src/thread/pthread_create.c +++ b/src/thread/pthread_create.c @@ -1,5 +1,6 @@ #include "pthread_impl.h" #include "stdio_impl.h" +#include static void dummy_0() { @@ -11,7 +12,6 @@ weak_alias(dummy_0, __pthread_tsd_run_dtors); _Noreturn void pthread_exit(void *result) { pthread_t self = pthread_self(); - int n; self->result = result; @@ -31,15 +31,26 @@ _Noreturn void pthread_exit(void *result) self->dead = 1; __unlock(self->killlock); - do n = libc.threads_minus_1; - while (n && a_cas(&libc.threads_minus_1, n, n-1)!=n); - if (!n) exit(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. */ + __syscall(SYS_rt_sigprocmask, SIG_BLOCK, SIGALL_SET, 0, _NSIG/8); + + if (a_fetch_add(&libc.threads_minus_1, -1)==0) exit(0); if (self->detached && self->map_base) { - if (self->detached == 2) - __syscall(SYS_set_tid_address, 0); - __syscall(SYS_rt_sigprocmask, SIG_BLOCK, - SIGALL_SET, 0, __SYSCALL_SSLEN); + /* 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); + + /* The following call unmaps the thread's stack mapping + * and then exits without touching the stack. */ __unmapself(self->map_base, self->map_size); } @@ -61,9 +72,18 @@ void __do_cleanup_pop(struct __ptcb *cb) static int start(void *p) { pthread_t self = p; + if (self->startlock[0]) { + __wait(self->startlock, 0, 1, 1); + if (self->startlock[0]) { + self->detached = 2; + pthread_exit(0); + } + __syscall(SYS_rt_sigprocmask, SIG_SETMASK, + self->sigmask, 0, _NSIG/8); + } if (self->unblock_cancel) __syscall(SYS_rt_sigprocmask, SIG_UNBLOCK, - SIGPT_SET, 0, __SYSCALL_SSLEN); + SIGPT_SET, 0, _NSIG/8); pthread_exit(self->start(self->start_arg)); return 0; } @@ -86,14 +106,15 @@ static void init_file_lock(FILE *f) void *__copy_tls(unsigned char *); -int pthread_create(pthread_t *restrict res, const pthread_attr_t *restrict attr, void *(*entry)(void *), void *restrict 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; + size_t size, guard; struct pthread *self = pthread_self(), *new; - unsigned char *map, *stack, *tsd; + unsigned char *map = 0, *stack = 0, *tsd = 0, *stack_limit; unsigned flags = 0x7d8f00; + int do_sched = 0; + pthread_attr_t attr = {0}; if (!self) return ENOSYS; if (!libc.threaded) { @@ -104,45 +125,70 @@ int pthread_create(pthread_t *restrict res, const pthread_attr_t *restrict attr, init_file_lock(__stderr_used); libc.threaded = 1; } + if (attrp) attr = *attrp; __acquire_ptc(); - if (attr && attr->_a_stackaddr) { - map = 0; - tsd = (void *)(attr->_a_stackaddr-__pthread_tsd_size & -16); - } else { - if (attr) { - guard = ROUND(attr->_a_guardsize + DEFAULT_GUARD_SIZE); - size = guard + ROUND(attr->_a_stacksize - + DEFAULT_STACK_SIZE + libc.tls_size); + 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; + } else { + size = ROUND(need); + guard = 0; } - size += __pthread_tsd_size; + } else { + guard = ROUND(DEFAULT_GUARD_SIZE + attr._a_guardsize); + size = guard + ROUND(DEFAULT_STACK_SIZE + 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) return EAGAIN; + if (map == MAP_FAILED) goto fail; if (mprotect(map+guard, size-guard, PROT_READ|PROT_WRITE)) { munmap(map, size); - return EAGAIN; + goto fail; } } else { map = mmap(0, size, PROT_READ|PROT_WRITE, MAP_PRIVATE|MAP_ANON, -1, 0); - if (map == MAP_FAILED) return EAGAIN; + if (map == MAP_FAILED) goto fail; } tsd = map + size - __pthread_tsd_size; + if (!stack) { + stack = tsd - libc.tls_size; + stack_limit = map + guard; + } } - new = __copy_tls(stack = tsd - libc.tls_size); + + new = __copy_tls(tsd - libc.tls_size); new->map_base = map; new->map_size = size; + new->stack = stack; + new->stack_size = stack - stack_limit; new->pid = self->pid; new->errno_ptr = &new->errno_val; new->start = entry; new->start_arg = arg; new->self = new; new->tsd = (void *)tsd; - if (attr && attr->_a_detach) { + if (attr._a_detach) { new->detached = 1; flags -= 0x200000; } + if (attr._a_sched) { + do_sched = new->startlock[0] = 1; + __syscall(SYS_rt_sigprocmask, SIG_BLOCK, + SIGALL_SET, self->sigmask, _NSIG/8); + } new->unblock_cancel = self->cancel; new->canary = self->canary; @@ -151,11 +197,28 @@ int pthread_create(pthread_t *restrict res, const pthread_attr_t *restrict attr, __release_ptc(); + if (do_sched) { + __syscall(SYS_rt_sigprocmask, SIG_SETMASK, + new->sigmask, 0, _NSIG/8); + } + 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; }