X-Git-Url: http://repo.macrolet.net/gitweb/?a=blobdiff_plain;f=src%2Fruntime%2Fthread.c;h=2c3367c3f7710d54eea6055083591fb51e5f7934;hb=8e9908b6f79b7bb52a281cfcbf0712d4b914f3cf;hp=117e3ec1bcb49a6a5a104dfc3779bd3cd4271630;hpb=862c0325616a991a5bd7b50d79f7176d2115493b;p=sbcl.git diff --git a/src/runtime/thread.c b/src/runtime/thread.c index 117e3ec..2c3367c 100644 --- a/src/runtime/thread.c +++ b/src/runtime/thread.c @@ -1,14 +1,36 @@ +/* + * This software is part of the SBCL system. See the README file for + * more information. + * + * This software is derived from the CMU CL system, which was + * written at Carnegie Mellon University and released into the + * public domain. The software is in the public domain and is + * provided with absolutely no warranty. See the COPYING and CREDITS + * files for more information. + */ + +#include "sbcl.h" + #include #include #include +#ifndef LISP_FEATURE_WIN32 #include +#endif #include #include #include #include +#ifndef LISP_FEATURE_WIN32 #include +#endif + +#ifdef LISP_FEATURE_MACH_EXCEPTION_HANDLER +#include +#include +#include +#endif -#include "sbcl.h" #include "runtime.h" #include "validate.h" /* for CONTROL_STACK_SIZE etc */ #include "alloc.h" @@ -23,60 +45,79 @@ #include "interr.h" /* for lose() */ #include "gc-internal.h" +#ifdef LISP_FEATURE_WIN32 +/* + * Win32 doesn't have SIGSTKSZ, and we're not switching stacks anyway, + * so define it arbitrarily + */ +#define SIGSTKSZ 1024 +#endif + +#if defined(LISP_FEATURE_DARWIN) && defined(LISP_FEATURE_SB_THREAD) +#define QUEUE_FREEABLE_THREAD_STACKS +#endif + +#ifdef LISP_FEATURE_FREEBSD +#define CREATE_CLEANUP_THREAD +#define LOCK_CREATE_THREAD +#endif + #define ALIEN_STACK_SIZE (1*1024*1024) /* 1Mb size chosen at random */ +struct freeable_stack { +#ifdef QUEUE_FREEABLE_THREAD_STACKS + struct freeable_stack *next; +#endif + os_thread_t os_thread; + os_vm_address_t stack; +}; + + +#ifdef QUEUE_FREEABLE_THREAD_STACKS +static struct freeable_stack * volatile freeable_stack_queue = 0; +static int freeable_stack_count = 0; +pthread_mutex_t freeable_stack_lock = PTHREAD_MUTEX_INITIALIZER; +#else +static struct freeable_stack * volatile freeable_stack = 0; +#endif + int dynamic_values_bytes=4096*sizeof(lispobj); /* same for all threads */ struct thread * volatile all_threads; extern struct interrupt_data * global_interrupt_data; -extern int linux_no_threads_p; #ifdef LISP_FEATURE_SB_THREAD - pthread_mutex_t all_threads_lock = PTHREAD_MUTEX_INITIALIZER; - -/* When trying to get all_threads_lock one should make sure that - * sig_stop_for_gc is not blocked. Else there would be a possible - * deadlock: gc locks it, other thread blocks signals, gc sends stop - * request to other thread and waits, other thread blocks on lock. */ -void check_sig_stop_for_gc_can_arrive_or_lose() -{ - /* Get the current sigmask, by blocking the empty set. */ - sigset_t empty,current; - sigemptyset(&empty); - thread_sigmask(SIG_BLOCK, &empty, ¤t); - if (sigismember(¤t,SIG_STOP_FOR_GC)) - lose("SIG_STOP_FOR_GC cannot arrive: it is blocked\n"); - if (SymbolValue(GC_INHIBIT,arch_os_get_current_thread()) != NIL) - lose("SIG_STOP_FOR_GC cannot arrive: gc is inhibited\n"); - if (arch_pseudo_atomic_atomic(NULL)) - lose("SIG_STOP_FOR_GC cannot arrive: in pseudo atomic\n"); -} - -#define GET_ALL_THREADS_LOCK(name) \ - { \ - sigset_t _newset,_oldset; \ - sigemptyset(&_newset); \ - sigaddset_deferrable(&_newset); \ - thread_sigmask(SIG_BLOCK, &_newset, &_oldset); \ - check_sig_stop_for_gc_can_arrive_or_lose(); \ - FSHOW_SIGNAL((stderr,"/%s:waiting on lock=%ld, thread=%lu\n",name, \ - all_threads_lock,arch_os_get_current_thread()->os_thread)); \ - pthread_mutex_lock(&all_threads_lock); \ - FSHOW_SIGNAL((stderr,"/%s:got lock, thread=%lu\n", \ - name,arch_os_get_current_thread()->os_thread)); - -#define RELEASE_ALL_THREADS_LOCK(name) \ - FSHOW_SIGNAL((stderr,"/%s:released lock\n",name)); \ - pthread_mutex_unlock(&all_threads_lock); \ - thread_sigmask(SIG_SETMASK,&_oldset,0); \ - } +#ifdef LOCK_CREATE_THREAD +static pthread_mutex_t create_thread_lock = PTHREAD_MUTEX_INITIALIZER; +#endif #endif - #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64) extern lispobj call_into_lisp_first_time(lispobj fun, lispobj *args, int nargs); #endif +static void +link_thread(struct thread *th) +{ + if (all_threads) all_threads->prev=th; + th->next=all_threads; + th->prev=0; + all_threads=th; +} + +#ifdef LISP_FEATURE_SB_THREAD +static void +unlink_thread(struct thread *th) +{ + if (th->prev) + th->prev->next = th->next; + else + all_threads = th->next; + if (th->next) + th->next->prev = th->prev; +} +#endif + static int initial_thread_trampoline(struct thread *th) { @@ -87,9 +128,12 @@ initial_thread_trampoline(struct thread *th) function = th->no_tls_value_marker; th->no_tls_value_marker = NO_TLS_VALUE_MARKER_WIDETAG; if(arch_os_thread_init(th)==0) return 1; + link_thread(th); + th->os_thread=thread_self(); +#ifndef LISP_FEATURE_WIN32 + protect_control_stack_guard_page(1); +#endif - if(th->os_thread < 1) lose("th->os_thread not set up right"); - th->state=STATE_RUNNING; #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64) return call_into_lisp_first_time(function,args,0); #else @@ -97,8 +141,133 @@ initial_thread_trampoline(struct thread *th) #endif } +#define THREAD_STRUCT_SIZE (THREAD_CONTROL_STACK_SIZE + BINDING_STACK_SIZE + \ + ALIEN_STACK_SIZE + dynamic_values_bytes + \ + 32 * SIGSTKSZ) + #ifdef LISP_FEATURE_SB_THREAD +#ifdef QUEUE_FREEABLE_THREAD_STACKS + +static void +queue_freeable_thread_stack(struct thread *thread_to_be_cleaned_up) +{ + if (thread_to_be_cleaned_up) { + pthread_mutex_lock(&freeable_stack_lock); + if (freeable_stack_queue) { + struct freeable_stack *new_freeable_stack = 0, *next; + next = freeable_stack_queue; + while (next->next) { + next = next->next; + } + new_freeable_stack = (struct freeable_stack *) + os_validate(0, sizeof(struct freeable_stack)); + new_freeable_stack->next = NULL; + new_freeable_stack->os_thread = thread_to_be_cleaned_up->os_thread; + new_freeable_stack->stack = (os_vm_address_t) + thread_to_be_cleaned_up->control_stack_start; + next->next = new_freeable_stack; + freeable_stack_count++; + } else { + struct freeable_stack *new_freeable_stack = 0; + new_freeable_stack = (struct freeable_stack *) + os_validate(0, sizeof(struct freeable_stack)); + new_freeable_stack->next = NULL; + new_freeable_stack->os_thread = thread_to_be_cleaned_up->os_thread; + new_freeable_stack->stack = (os_vm_address_t) + thread_to_be_cleaned_up->control_stack_start; + freeable_stack_queue = new_freeable_stack; + freeable_stack_count++; + } + pthread_mutex_unlock(&freeable_stack_lock); + } +} + +#define FREEABLE_STACK_QUEUE_SIZE 4 + +static void +free_freeable_stacks() { + if (freeable_stack_queue && (freeable_stack_count > FREEABLE_STACK_QUEUE_SIZE)) { + struct freeable_stack* old; + pthread_mutex_lock(&freeable_stack_lock); + old = freeable_stack_queue; + freeable_stack_queue = old->next; + freeable_stack_count--; + gc_assert(pthread_join(old->os_thread, NULL) == 0); + FSHOW((stderr, "freeing thread %x stack\n", old->os_thread)); + os_invalidate(old->stack, THREAD_STRUCT_SIZE); + os_invalidate((os_vm_address_t)old, sizeof(struct freeable_stack)); + pthread_mutex_unlock(&freeable_stack_lock); + } +} + +#elif defined(CREATE_CLEANUP_THREAD) +static void * +cleanup_thread(void *arg) +{ + struct freeable_stack *freeable = arg; + pthread_t self = pthread_self(); + + FSHOW((stderr, "/cleaner thread(%p): joining %p\n", + self, freeable->os_thread)); + gc_assert(pthread_join(freeable->os_thread, NULL) == 0); + FSHOW((stderr, "/cleaner thread(%p): free stack %p\n", + self, freeable->stack)); + os_invalidate(freeable->stack, THREAD_STRUCT_SIZE); + free(freeable); + + pthread_detach(self); + + return NULL; +} + +static void +create_cleanup_thread(struct thread *thread_to_be_cleaned_up) +{ + pthread_t thread; + int result; + + if (thread_to_be_cleaned_up) { + struct freeable_stack *freeable = + malloc(sizeof(struct freeable_stack)); + gc_assert(freeable != NULL); + freeable->os_thread = thread_to_be_cleaned_up->os_thread; + freeable->stack = + (os_vm_address_t) thread_to_be_cleaned_up->control_stack_start; + result = pthread_create(&thread, NULL, cleanup_thread, freeable); + gc_assert(result == 0); + sched_yield(); + } +} + +#else +static void +free_thread_stack_later(struct thread *thread_to_be_cleaned_up) +{ + struct freeable_stack *new_freeable_stack = 0; + if (thread_to_be_cleaned_up) { + new_freeable_stack = (struct freeable_stack *) + os_validate(0, sizeof(struct freeable_stack)); + new_freeable_stack->os_thread = thread_to_be_cleaned_up->os_thread; + new_freeable_stack->stack = (os_vm_address_t) + thread_to_be_cleaned_up->control_stack_start; + } + new_freeable_stack = (struct freeable_stack *) + swap_lispobjs((lispobj *)(void *)&freeable_stack, + (lispobj)new_freeable_stack); + if (new_freeable_stack) { + FSHOW((stderr,"/reaping %p\n", (void*) new_freeable_stack->os_thread)); + /* Under NPTL pthread_join really waits until the thread + * exists and the stack can be safely freed. This is sadly not + * mandated by the pthread spec. */ + gc_assert(pthread_join(new_freeable_stack->os_thread, NULL) == 0); + os_invalidate(new_freeable_stack->stack, THREAD_STRUCT_SIZE); + os_invalidate((os_vm_address_t) new_freeable_stack, + sizeof(struct freeable_stack)); + } +} +#endif + /* this is the first thing that runs in the child (which is why the * silly calling convention). Basically it calls the user's requested * lisp function after doing arch_os_thread_init and whatever other @@ -108,7 +277,9 @@ int new_thread_trampoline(struct thread *th) { lispobj function; - int result; + int result, lock_ret; + + FSHOW((stderr,"/creating thread %lu\n", thread_self())); function = th->no_tls_value_marker; th->no_tls_value_marker = NO_TLS_VALUE_MARKER_WIDETAG; if(arch_os_thread_init(th)==0) { @@ -116,22 +287,62 @@ new_thread_trampoline(struct thread *th) lose("arch_os_thread_init failed\n"); } - /* wait here until our thread is linked into all_threads: see below */ - { - volatile os_thread_t *tid=&th->os_thread; - while(*tid<1) sched_yield(); - } + th->os_thread=thread_self(); + protect_control_stack_guard_page(1); + /* Since GC can only know about this thread from the all_threads + * list and we're just adding this thread to it there is no danger + * of deadlocking even with SIG_STOP_FOR_GC blocked (which it is + * not). */ + lock_ret = pthread_mutex_lock(&all_threads_lock); + gc_assert(lock_ret == 0); + link_thread(th); + lock_ret = pthread_mutex_unlock(&all_threads_lock); + gc_assert(lock_ret == 0); - th->state=STATE_RUNNING; result = funcall0(function); + + /* Block GC */ + block_blockable_signals(); th->state=STATE_DEAD; + + /* SIG_STOP_FOR_GC is blocked and GC might be waiting for this + * thread, but since we are already dead it won't wait long. */ + lock_ret = pthread_mutex_lock(&all_threads_lock); + gc_assert(lock_ret == 0); + + gc_alloc_update_page_tables(0, &th->alloc_region); + unlink_thread(th); + pthread_mutex_unlock(&all_threads_lock); + gc_assert(lock_ret == 0); + + if(th->tls_cookie>=0) arch_os_thread_cleanup(th); + os_invalidate((os_vm_address_t)th->interrupt_data, + (sizeof (struct interrupt_data))); + +#ifdef LISP_FEATURE_MACH_EXCEPTION_HANDLER + FSHOW((stderr, "Deallocating mach port %x\n", THREAD_STRUCT_TO_EXCEPTION_PORT(th))); + mach_port_move_member(mach_task_self(), + THREAD_STRUCT_TO_EXCEPTION_PORT(th), + MACH_PORT_NULL); + mach_port_deallocate(mach_task_self(), + THREAD_STRUCT_TO_EXCEPTION_PORT(th)); + mach_port_destroy(mach_task_self(), + THREAD_STRUCT_TO_EXCEPTION_PORT(th)); +#endif + +#ifdef QUEUE_FREEABLE_THREAD_STACKS + queue_freeable_thread_stack(th); +#elif defined(CREATE_CLEANUP_THREAD) + create_cleanup_thread(th); +#else + free_thread_stack_later(th); +#endif + + FSHOW((stderr,"/exiting thread %p\n", thread_self())); return result; } -#endif /* LISP_FEATURE_SB_THREAD */ -#define THREAD_STRUCT_SIZE (THREAD_CONTROL_STACK_SIZE + BINDING_STACK_SIZE + \ - ALIEN_STACK_SIZE + dynamic_values_bytes + \ - 32 * SIGSTKSZ) +#endif /* LISP_FEATURE_SB_THREAD */ static void free_thread_struct(struct thread *th) @@ -153,7 +364,14 @@ create_thread_struct(lispobj initial_function) { union per_thread_data *per_thread; struct thread *th=0; /* subdue gcc */ void *spaces=0; +#ifdef LISP_FEATURE_SB_THREAD + int i; +#endif +#ifdef CREATE_CLEANUP_THREAD + /* Give a chance for cleanup threads to run. */ + sched_yield(); +#endif /* may as well allocate all the spaces at once: it saves us from * having to decide what to do if only some of the allocations * succeed */ @@ -166,20 +384,19 @@ create_thread_struct(lispobj initial_function) { BINDING_STACK_SIZE+ ALIEN_STACK_SIZE); - if(all_threads) { - memcpy(per_thread,arch_os_get_current_thread(), - dynamic_values_bytes); - } else { #ifdef LISP_FEATURE_SB_THREAD - int i; - for(i=0;i<(dynamic_values_bytes/sizeof(lispobj));i++) - per_thread->dynamic_values[i]=NO_TLS_VALUE_MARKER_WIDETAG; - if(SymbolValue(FREE_TLS_INDEX,0)==UNBOUND_MARKER_WIDETAG) + for(i = 0; i < (dynamic_values_bytes / sizeof(lispobj)); i++) + per_thread->dynamic_values[i] = NO_TLS_VALUE_MARKER_WIDETAG; + if (all_threads == 0) { + if(SymbolValue(FREE_TLS_INDEX,0)==UNBOUND_MARKER_WIDETAG) { SetSymbolValue (FREE_TLS_INDEX, + /* FIXME: should be MAX_INTERRUPTS -1 ? */ make_fixnum(MAX_INTERRUPTS+ sizeof(struct thread)/sizeof(lispobj)), 0); + SetSymbolValue(TLS_INDEX_LOCK,make_fixnum(0),0); + } #define STATIC_TLS_INIT(sym,field) \ ((struct symbol *)(sym-OTHER_POINTER_LOWTAG))->tls_index= \ make_fixnum(THREAD_SLOT_OFFSET_WORDS(field)) @@ -190,12 +407,11 @@ create_thread_struct(lispobj initial_function) { STATIC_TLS_INIT(CONTROL_STACK_END,control_stack_end); STATIC_TLS_INIT(ALIEN_STACK,alien_stack_pointer); #if defined(LISP_FEATURE_X86) || defined (LISP_FEATURE_X86_64) - STATIC_TLS_INIT(PSEUDO_ATOMIC_ATOMIC,pseudo_atomic_atomic); - STATIC_TLS_INIT(PSEUDO_ATOMIC_INTERRUPTED,pseudo_atomic_interrupted); + STATIC_TLS_INIT(PSEUDO_ATOMIC_BITS,pseudo_atomic_bits); #endif #undef STATIC_TLS_INIT -#endif } +#endif th=&per_thread->thread; th->control_stack_start = spaces; @@ -207,9 +423,7 @@ create_thread_struct(lispobj initial_function) { th->binding_stack_pointer=th->binding_stack_start; th->this=th; th->os_thread=0; - th->interrupt_fun=NIL; - th->interrupt_fun_lock=0; - th->state=STATE_STARTING; + th->state=STATE_RUNNING; #ifdef LISP_FEATURE_STACK_GROWS_DOWNWARD_NOT_UPWARD th->alien_stack_pointer=((void *)th->alien_stack_start + ALIEN_STACK_SIZE-N_WORD_BYTES); @@ -217,8 +431,7 @@ create_thread_struct(lispobj initial_function) { th->alien_stack_pointer=((void *)th->alien_stack_start); #endif #if defined(LISP_FEATURE_X86) || defined (LISP_FEATURE_X86_64) - th->pseudo_atomic_interrupted=0; - th->pseudo_atomic_atomic=0; + th->pseudo_atomic_bits=0; #endif #ifdef LISP_FEATURE_GENCGC gc_set_region_empty(&th->alloc_region); @@ -237,8 +450,7 @@ create_thread_struct(lispobj initial_function) { #if defined(LISP_FEATURE_X86) || defined (LISP_FEATURE_X86_64) SetSymbolValue(BINDING_STACK_POINTER,(lispobj)th->binding_stack_pointer,th); SetSymbolValue(ALIEN_STACK,(lispobj)th->alien_stack_pointer,th); - SetSymbolValue(PSEUDO_ATOMIC_ATOMIC,(lispobj)th->pseudo_atomic_atomic,th); - SetSymbolValue(PSEUDO_ATOMIC_INTERRUPTED,th->pseudo_atomic_interrupted,th); + SetSymbolValue(PSEUDO_ATOMIC_BITS,(lispobj)th->pseudo_atomic_bits,th); #else current_binding_stack_pointer=th->binding_stack_pointer; current_control_stack_pointer=th->control_stack_start; @@ -260,42 +472,29 @@ create_thread_struct(lispobj initial_function) { free_thread_struct(th); return 0; } - if(all_threads) - memcpy(th->interrupt_data, - arch_os_get_current_thread()->interrupt_data, - sizeof (struct interrupt_data)); - else - memcpy(th->interrupt_data,global_interrupt_data, - sizeof (struct interrupt_data)); - + th->interrupt_data->pending_handler = 0; th->no_tls_value_marker=initial_function; + + th->stepping = NIL; return th; } -static void -link_thread(struct thread *th,os_thread_t kid_tid) -{ - if (all_threads) all_threads->prev=th; - th->next=all_threads; - th->prev=0; - all_threads=th; - /* note that th->os_thread is 0 at this time. We rely on - * all_threads_lock to ensure that we don't have >1 thread with - * os_thread=0 on the list at once - */ - protect_control_stack_guard_page(th,1); - /* child will not start until this is set */ - th->os_thread=kid_tid; - FSHOW((stderr,"/created thread %lu\n",kid_tid)); -} +#ifdef LISP_FEATURE_MACH_EXCEPTION_HANDLER +mach_port_t setup_mach_exception_handling_thread(); +kern_return_t mach_thread_init(mach_port_t thread_exception_port); + +#endif void create_initial_thread(lispobj initial_function) { struct thread *th=create_thread_struct(initial_function); - os_thread_t kid_tid=thread_self(); - if(th && kid_tid>0) { - link_thread(th,kid_tid); - initial_thread_trampoline(all_threads); /* no return */ - } else lose("can't create initial thread"); + if(th) { +#ifdef LISP_FEATURE_MACH_EXCEPTION_HANDLER + kern_return_t ret; + + setup_mach_exception_handling_thread(); +#endif + initial_thread_trampoline(th); /* no return */ + } else lose("can't create initial thread\n"); } #ifdef LISP_FEATURE_SB_THREAD @@ -312,83 +511,80 @@ boolean create_os_thread(struct thread *th,os_thread_t *kid_tid) pthread_attr_t attr; sigset_t newset,oldset; boolean r=1; + int retcode, initcode, sizecode, addrcode; + + FSHOW_SIGNAL((stderr,"/create_os_thread: creating new thread\n")); + +#ifdef LOCK_CREATE_THREAD + retcode = pthread_mutex_lock(&create_thread_lock); + gc_assert(retcode == 0); + FSHOW_SIGNAL((stderr,"/create_os_thread: got lock\n")); +#endif sigemptyset(&newset); - /* Blocking deferrable signals is enough, since gc_stop_the_world - * waits until the child leaves STATE_STARTING. And why not let gc - * proceed as soon as possible? */ + /* Blocking deferrable signals is enough, no need to block + * SIG_STOP_FOR_GC because the child process is not linked onto + * all_threads until it's ready. */ sigaddset_deferrable(&newset); thread_sigmask(SIG_BLOCK, &newset, &oldset); - if((pthread_attr_init(&attr)) || +#if defined(LISP_FEATURE_DARWIN) +#define CONTROL_STACK_ADJUST 8192 /* darwin wants page-aligned stacks */ +#else +#define CONTROL_STACK_ADJUST 16 +#endif + + if((initcode = pthread_attr_init(&attr)) || + /* FIXME: why do we even have this in the first place? */ (pthread_attr_setstack(&attr,th->control_stack_start, - THREAD_CONTROL_STACK_SIZE-16)) || - (pthread_create - (kid_tid,&attr,(void *(*)(void *))new_thread_trampoline,th))) + THREAD_CONTROL_STACK_SIZE-CONTROL_STACK_ADJUST)) || +#undef CONTROL_STACK_ADJUST + (retcode = pthread_create + (kid_tid,&attr,(void *(*)(void *))new_thread_trampoline,th))) { + FSHOW_SIGNAL((stderr, "init, size, addr = %d, %d, %d\n", initcode, sizecode, addrcode)); + FSHOW_SIGNAL((stderr, printf("pthread_create returned %d, errno %d\n", retcode, errno))); + FSHOW_SIGNAL((stderr, "wanted stack size %d, min stack size %d\n", + THREAD_CONTROL_STACK_SIZE-16, PTHREAD_STACK_MIN)); + if(retcode < 0) { + perror("create_os_thread"); + } r=0; + } + +#ifdef QUEUE_FREEABLE_THREAD_STACKS + free_freeable_stacks(); +#endif thread_sigmask(SIG_SETMASK,&oldset,0); +#ifdef LOCK_CREATE_THREAD + retcode = pthread_mutex_unlock(&create_thread_lock); + gc_assert(retcode == 0); + FSHOW_SIGNAL((stderr,"/create_os_thread: released lock\n")); +#endif return r; } -struct thread *create_thread(lispobj initial_function) { +os_thread_t create_thread(lispobj initial_function) { struct thread *th; - os_thread_t kid_tid=0; - boolean success; - - if(linux_no_threads_p) return 0; + os_thread_t kid_tid; - th=create_thread_struct(initial_function); + /* Assuming that a fresh thread struct has no lisp objects in it, + * linking it to all_threads can be left to the thread itself + * without fear of gc lossage. initial_function violates this + * assumption and must stay pinned until the child starts up. */ + th = create_thread_struct(initial_function); if(th==0) return 0; - /* we must not be interrupted here after a successful - * create_os_thread, because the kid will be waiting for its - * thread struct to be linked */ - GET_ALL_THREADS_LOCK("create_thread") - - success=create_os_thread(th,&kid_tid); - if (success) - link_thread(th,kid_tid); - else + if (create_os_thread(th,&kid_tid)) { + return kid_tid; + } else { free_thread_struct(th); - - RELEASE_ALL_THREADS_LOCK("create_thread") - - if (success) - return th; - else return 0; -} - -/* called from lisp from the thread object finalizer */ -void reap_dead_thread(struct thread *th) -{ - if(th->state!=STATE_DEAD) - lose("thread %p is not joinable, state=%d\n",th,th->state); -#ifdef LISP_FEATURE_GENCGC - { - sigset_t newset,oldset; - sigemptyset(&newset); - sigaddset_blockable(&newset); - thread_sigmask(SIG_BLOCK, &newset, &oldset); - gc_alloc_update_page_tables(0, &th->alloc_region); - thread_sigmask(SIG_SETMASK,&oldset,0); } -#endif - GET_ALL_THREADS_LOCK("reap_dead_thread") - FSHOW((stderr,"/reap_dead_thread: reaping %lu\n",th->os_thread)); - if(th->prev) - th->prev->next=th->next; - else all_threads=th->next; - if(th->next) - th->next->prev=th->prev; - RELEASE_ALL_THREADS_LOCK("reap_dead_thread") - if(th->tls_cookie>=0) arch_os_thread_cleanup(th); - gc_assert(pthread_join(th->os_thread,NULL)==0); - free_thread_struct(th); } /* Send the signo to os_thread, retry if the rt signal queue is * full. */ -static int kill_thread_safely(os_thread_t os_thread, int signo) +int +kill_thread_safely(os_thread_t os_thread, int signo) { int r; /* The man page does not mention EAGAIN as a valid return value @@ -407,56 +603,16 @@ static int kill_thread_safely(os_thread_t os_thread, int signo) return r; } -int interrupt_thread(struct thread *th, lispobj function) +int signal_interrupt_thread(os_thread_t os_thread) { - /* In clone_threads, if A and B both interrupt C at approximately - * the same time, it does not matter: the second signal will be - * masked until the handler has returned from the first one. In - * pthreads though, we can't put the knowledge of what function to - * call into the siginfo, so we have to store it in the - * destination thread, and do it in such a way that A won't - * clobber B's interrupt. Hence, this stupid linked list. - * - * This does depend on SIG_INTERRUPT_THREAD being queued (as POSIX - * RT signals are): we need to keep interrupt_fun data for exactly - * as many signals as are going to be received by the destination - * thread. - */ - lispobj c=alloc_cons(function,NIL); - sigset_t newset,oldset; - sigemptyset(&newset); - /* interrupt_thread_handler locks this spinlock with blockables - * blocked (it does so for the sake of - * arrange_return_to_lisp_function), so we must also block them or - * else SIG_STOP_FOR_GC and all_threads_lock will find a way to - * deadlock. */ - sigaddset_blockable(&newset); - thread_sigmask(SIG_BLOCK, &newset, &oldset); - if (th == arch_os_get_current_thread()) - lose("cannot interrupt current thread"); - get_spinlock(&th->interrupt_fun_lock, - (long)arch_os_get_current_thread()); - ((struct cons *)native_pointer(c))->cdr=th->interrupt_fun; - th->interrupt_fun=c; - release_spinlock(&th->interrupt_fun_lock); - thread_sigmask(SIG_SETMASK,&oldset,0); - /* Called from lisp with the thread object as a parameter. Thus, - * the object cannot be garbage collected and consequently reaped - * and joined. Because it's not joined, kill should work (even if - * the thread has died/exited). */ - { - int status=kill_thread_safely(th->os_thread,SIG_INTERRUPT_THREAD); - if (status==0) { - return 0; - } else if (status==ESRCH) { - /* This thread has exited. */ - th->interrupt_fun=NIL; - errno=ESRCH; - return -1; - } else { - lose("cannot send SIG_INTERRUPT_THREAD to thread=%lu: %d, %s", - th->os_thread,status,strerror(status)); - } + int status = kill_thread_safely(os_thread, SIG_INTERRUPT_THREAD); + if (status == 0) { + return 0; + } else if (status == ESRCH) { + return -1; + } else { + lose("cannot send SIG_INTERRUPT_THREAD to thread=%lu: %d, %s\n", + os_thread, status, strerror(status)); } } @@ -472,25 +628,38 @@ int interrupt_thread(struct thread *th, lispobj function) void gc_stop_the_world() { struct thread *p,*th=arch_os_get_current_thread(); - int status; + int status, lock_ret; +#ifdef LOCK_CREATE_THREAD + /* KLUDGE: Stopping the thread during pthread_create() causes deadlock + * on FreeBSD. */ + FSHOW_SIGNAL((stderr,"/gc_stop_the_world:waiting on create_thread_lock, thread=%lu\n", + th->os_thread)); + lock_ret = pthread_mutex_lock(&create_thread_lock); + gc_assert(lock_ret == 0); + FSHOW_SIGNAL((stderr,"/gc_stop_the_world:got create_thread_lock, thread=%lu\n", + th->os_thread)); +#endif FSHOW_SIGNAL((stderr,"/gc_stop_the_world:waiting on lock, thread=%lu\n", th->os_thread)); /* keep threads from starting while the world is stopped. */ - pthread_mutex_lock(&all_threads_lock); \ + lock_ret = pthread_mutex_lock(&all_threads_lock); \ + gc_assert(lock_ret == 0); + FSHOW_SIGNAL((stderr,"/gc_stop_the_world:got lock, thread=%lu\n", th->os_thread)); /* stop all other threads by sending them SIG_STOP_FOR_GC */ for(p=all_threads; p; p=p->next) { - while(p->state==STATE_STARTING) sched_yield(); - if((p!=th) && (p->state==STATE_RUNNING)) { - FSHOW_SIGNAL((stderr,"/gc_stop_the_world: suspending %lu\n", - p->os_thread)); + gc_assert(p->os_thread != 0); + FSHOW_SIGNAL((stderr,"/gc_stop_the_world: p->state: %x\n", p->state)); + if((p!=th) && ((p->state==STATE_RUNNING))) { + FSHOW_SIGNAL((stderr,"/gc_stop_the_world: suspending %x, os_thread %x\n", + p, p->os_thread)); status=kill_thread_safely(p->os_thread,SIG_STOP_FOR_GC); if (status==ESRCH) { /* This thread has exited. */ gc_assert(p->state==STATE_DEAD); } else if (status) { - lose("cannot send suspend thread=%lu: %d, %s", + lose("cannot send suspend thread=%lu: %d, %s\n", p->os_thread,status,strerror(status)); } } @@ -498,13 +667,11 @@ void gc_stop_the_world() FSHOW_SIGNAL((stderr,"/gc_stop_the_world:signals sent\n")); /* wait for the running threads to stop or finish */ for(p=all_threads;p;) { - gc_assert(p->os_thread!=0); - gc_assert(p->state!=STATE_STARTING); - if((p==th) || (p->state==STATE_SUSPENDED) || - (p->state==STATE_DEAD)) { - p=p->next; - } else { + FSHOW_SIGNAL((stderr,"/gc_stop_the_world: th: %p, p: %p\n", th, p)); + if((p!=th) && (p->state==STATE_RUNNING)) { sched_yield(); + } else { + p=p->next; } } FSHOW_SIGNAL((stderr,"/gc_stop_the_world:end\n")); @@ -513,7 +680,7 @@ void gc_stop_the_world() void gc_start_the_world() { struct thread *p,*th=arch_os_get_current_thread(); - int status; + int status, lock_ret; /* if a resumed thread creates a new thread before we're done with * this loop, the new thread will get consed on the front of * all_threads, but it won't have been stopped so won't need @@ -529,9 +696,14 @@ void gc_start_the_world() FSHOW_SIGNAL((stderr, "/gc_start_the_world: resuming %lu\n", p->os_thread)); p->state=STATE_RUNNING; + +#if defined(SIG_RESUME_FROM_GC) + status=kill_thread_safely(p->os_thread,SIG_RESUME_FROM_GC); +#else status=kill_thread_safely(p->os_thread,SIG_STOP_FOR_GC); +#endif if (status) { - lose("cannot resume thread=%lu: %d, %s", + lose("cannot resume thread=%lu: %d, %s\n", p->os_thread,status,strerror(status)); } } @@ -540,7 +712,14 @@ void gc_start_the_world() * SIG_STOP_FOR_GC wouldn't need to be a rt signal. That has some * performance implications, but does away with the 'rt signal * queue full' problem. */ - pthread_mutex_unlock(&all_threads_lock); \ + + lock_ret = pthread_mutex_unlock(&all_threads_lock); + gc_assert(lock_ret == 0); +#ifdef LOCK_CREATE_THREAD + lock_ret = pthread_mutex_unlock(&create_thread_lock); + gc_assert(lock_ret == 0); +#endif + FSHOW_SIGNAL((stderr,"/gc_start_the_world:end\n")); } #endif