X-Git-Url: http://repo.macrolet.net/gitweb/?a=blobdiff_plain;f=src%2Fruntime%2Fthread.c;h=ab0db0886eab31ea383d5a85b60349a046e63e4d;hb=656f994cdddc89af3a99c8af266816b09879df4a;hp=a6b462fad6a6b61c31c978423b7441c3140b1d07;hpb=d8e682fdfb7e8ba067e15aea0f3d1f8d37ca9eb1;p=sbcl.git diff --git a/src/runtime/thread.c b/src/runtime/thread.c index a6b462f..ab0db08 100644 --- a/src/runtime/thread.c +++ b/src/runtime/thread.c @@ -10,7 +10,7 @@ #include "sbcl.h" #include "runtime.h" -#include "validate.h" /* for CONTROL_STACK_SIZE etc */ +#include "validate.h" /* for CONTROL_STACK_SIZE etc */ #include "alloc.h" #include "thread.h" #include "arch.h" @@ -25,23 +25,67 @@ #define ALIEN_STACK_SIZE (1*1024*1024) /* 1Mb size chosen at random */ -int dynamic_values_bytes=4096*sizeof(lispobj); /* same for all threads */ -struct thread *all_threads; -volatile lispobj all_threads_lock; -volatile lispobj thread_start_lock; +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; -int +#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); \ + } +#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 int initial_thread_trampoline(struct thread *th) { lispobj function; #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64) lispobj *args = NULL; #endif - - function = th->unbound_marker; - th->unbound_marker = UNBOUND_MARKER_WIDETAG; + 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; if(th->os_thread < 1) lose("th->os_thread not set up right"); @@ -54,14 +98,6 @@ initial_thread_trampoline(struct thread *th) } #ifdef LISP_FEATURE_SB_THREAD -void mark_thread_dead(struct thread *th) { - funcall1(SymbolFunction(HANDLE_THREAD_EXIT),alloc_number(th->os_thread)); - /* I hope it's safe for a thread to detach itself inside a - * cancellation cleanup */ - pthread_detach(th->os_thread); - th->state=STATE_DEAD; - /* FIXME: if gc hits here it will rip the stack from under us */ -} /* this is the first thing that runs in the child (which is why the * silly calling convention). Basically it calls the user's requested @@ -71,76 +107,93 @@ void mark_thread_dead(struct thread *th) { int new_thread_trampoline(struct thread *th) { - lispobj function,ret; - function = th->unbound_marker; - th->unbound_marker = UNBOUND_MARKER_WIDETAG; - pthread_cleanup_push((void (*) (void *))mark_thread_dead,th); - if(arch_os_thread_init(th)==0) return 1; + lispobj function; + int result; + function = th->no_tls_value_marker; + th->no_tls_value_marker = NO_TLS_VALUE_MARKER_WIDETAG; + if(arch_os_thread_init(th)==0) { + /* FIXME: handle error */ + lose("arch_os_thread_init failed\n"); + } /* wait here until our thread is linked into all_threads: see below */ - while(th->os_thread<1) sched_yield(); + { + volatile os_thread_t *tid=&th->os_thread; + while(*tid<1) sched_yield(); + } th->state=STATE_RUNNING; - ret = funcall0(function); - /* execute cleanup */ - pthread_cleanup_pop(1); - return ret; + result = funcall0(function); + th->state=STATE_DEAD; + 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) + +static void +free_thread_struct(struct thread *th) +{ + if (th->interrupt_data) + os_invalidate((os_vm_address_t) th->interrupt_data, + (sizeof (struct interrupt_data))); + os_invalidate((os_vm_address_t) th->control_stack_start, + THREAD_STRUCT_SIZE); +} + /* this is called from any other thread to create the new one, and - * initialize all parts of it that can be initialized from another - * thread + * initialize all parts of it that can be initialized from another + * thread */ -struct thread * create_thread_struct(lispobj initial_function) { +static struct thread * +create_thread_struct(lispobj initial_function) { union per_thread_data *per_thread; - struct thread *th=0; /* subdue gcc */ + struct thread *th=0; /* subdue gcc */ void *spaces=0; /* 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 */ - spaces=os_validate(0, - THREAD_CONTROL_STACK_SIZE+ - BINDING_STACK_SIZE+ - ALIEN_STACK_SIZE+ - dynamic_values_bytes+ - 32*SIGSTKSZ); + spaces=os_validate(0, THREAD_STRUCT_SIZE); if(!spaces) - return NULL; + return NULL; per_thread=(union per_thread_data *) - (spaces+ - THREAD_CONTROL_STACK_SIZE+ - BINDING_STACK_SIZE+ - ALIEN_STACK_SIZE); + (spaces+ + THREAD_CONTROL_STACK_SIZE+ + BINDING_STACK_SIZE+ + ALIEN_STACK_SIZE); if(all_threads) { - memcpy(per_thread,arch_os_get_current_thread(), - dynamic_values_bytes); + 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]=UNBOUND_MARKER_WIDETAG; - if(SymbolValue(FREE_TLS_INDEX,0)==UNBOUND_MARKER_WIDETAG) - SetSymbolValue - (FREE_TLS_INDEX, - make_fixnum(MAX_INTERRUPTS+ - sizeof(struct thread)/sizeof(lispobj)), - 0); + 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) { + SetSymbolValue + (FREE_TLS_INDEX, + 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)) - - STATIC_TLS_INIT(BINDING_STACK_START,binding_stack_start); - STATIC_TLS_INIT(BINDING_STACK_POINTER,binding_stack_pointer); - STATIC_TLS_INIT(CONTROL_STACK_START,control_stack_start); - STATIC_TLS_INIT(CONTROL_STACK_END,control_stack_end); - STATIC_TLS_INIT(ALIEN_STACK,alien_stack_pointer); -#ifdef LISP_FEATURE_X86 - STATIC_TLS_INIT(PSEUDO_ATOMIC_ATOMIC,pseudo_atomic_atomic); - STATIC_TLS_INIT(PSEUDO_ATOMIC_INTERRUPTED,pseudo_atomic_interrupted); + + STATIC_TLS_INIT(BINDING_STACK_START,binding_stack_start); + STATIC_TLS_INIT(BINDING_STACK_POINTER,binding_stack_pointer); + STATIC_TLS_INIT(CONTROL_STACK_START,control_stack_start); + 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); #endif #undef STATIC_TLS_INIT #endif @@ -149,10 +202,10 @@ struct thread * create_thread_struct(lispobj initial_function) { th=&per_thread->thread; th->control_stack_start = spaces; th->binding_stack_start= - (lispobj*)((void*)th->control_stack_start+THREAD_CONTROL_STACK_SIZE); + (lispobj*)((void*)th->control_stack_start+THREAD_CONTROL_STACK_SIZE); th->control_stack_end = th->binding_stack_start; th->alien_stack_start= - (lispobj*)((void*)th->binding_stack_start+BINDING_STACK_SIZE); + (lispobj*)((void*)th->binding_stack_start+BINDING_STACK_SIZE); th->binding_stack_pointer=th->binding_stack_start; th->this=th; th->os_thread=0; @@ -161,7 +214,7 @@ struct thread * create_thread_struct(lispobj initial_function) { th->state=STATE_STARTING; #ifdef LISP_FEATURE_STACK_GROWS_DOWNWARD_NOT_UPWARD th->alien_stack_pointer=((void *)th->alien_stack_start - + ALIEN_STACK_SIZE-N_WORD_BYTES); + + ALIEN_STACK_SIZE-N_WORD_BYTES); #else th->alien_stack_pointer=((void *)th->alien_stack_start); #endif @@ -176,7 +229,7 @@ struct thread * create_thread_struct(lispobj initial_function) { #ifndef LISP_FEATURE_SB_THREAD /* the tls-points-into-struct-thread trick is only good for threaded * sbcl, because unithread sbcl doesn't have tls. So, we copy the - * appropriate values from struct thread here, and make sure that + * appropriate values from struct thread here, and make sure that * we use the appropriate SymbolValue macros to access any of the * variable quantities from the C runtime. It's not quite OAOOM, * it just feels like it */ @@ -192,35 +245,31 @@ struct thread * create_thread_struct(lispobj initial_function) { current_binding_stack_pointer=th->binding_stack_pointer; current_control_stack_pointer=th->control_stack_start; #endif -#endif +#endif bind_variable(CURRENT_CATCH_BLOCK,make_fixnum(0),th); - bind_variable(CURRENT_UNWIND_PROTECT_BLOCK,make_fixnum(0),th); + bind_variable(CURRENT_UNWIND_PROTECT_BLOCK,make_fixnum(0),th); bind_variable(FREE_INTERRUPT_CONTEXT_INDEX,make_fixnum(0),th); bind_variable(INTERRUPT_PENDING, NIL,th); bind_variable(INTERRUPTS_ENABLED,T,th); + bind_variable(GC_PENDING,NIL,th); +#ifdef LISP_FEATURE_SB_THREAD + bind_variable(STOP_FOR_GC_PENDING,NIL,th); +#endif th->interrupt_data = (struct interrupt_data *) os_validate(0,(sizeof (struct interrupt_data))); - 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->unbound_marker=initial_function; + if (!th->interrupt_data) { + free_thread_struct(th); + return 0; + } + th->interrupt_data->pending_handler = 0; + th->no_tls_value_marker=initial_function; return th; } -void link_thread(struct thread *th,os_thread_t kid_tid) +static void +link_thread(struct thread *th,os_thread_t kid_tid) { - sigset_t newset,oldset; - sigemptyset(&newset); - sigaddset_blockable(&newset); - thread_sigmask(SIG_BLOCK, &newset, &oldset); - - get_spinlock(&all_threads_lock,kid_tid); if (all_threads) all_threads->prev=th; th->next=all_threads; th->prev=0; @@ -229,233 +278,264 @@ void link_thread(struct thread *th,os_thread_t kid_tid) * 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->os_thread,1); + protect_control_stack_guard_page(th,1); /* child will not start until this is set */ th->os_thread=kid_tid; - release_spinlock(&all_threads_lock); - - thread_sigmask(SIG_SETMASK,&oldset,0); + FSHOW((stderr,"/created thread %lu\n",kid_tid)); } 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 */ + link_thread(th,kid_tid); + initial_thread_trampoline(all_threads); /* no return */ } else lose("can't create initial thread"); } #ifdef LISP_FEATURE_SB_THREAD -os_thread_t create_thread(lispobj initial_function) { + +#ifndef __USE_XOPEN2K +extern int pthread_attr_setstack (pthread_attr_t *__attr, void *__stackaddr, + size_t __stacksize); +#endif + +boolean create_os_thread(struct thread *th,os_thread_t *kid_tid) +{ + /* The new thread inherits the restrictive signal mask set here, + * and enables signals again when it is set up properly. */ + pthread_attr_t attr; + sigset_t newset,oldset; + boolean r=1; + 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? */ + sigaddset_deferrable(&newset); + thread_sigmask(SIG_BLOCK, &newset, &oldset); + + if((pthread_attr_init(&attr)) || + (pthread_attr_setstack(&attr,th->control_stack_start, + THREAD_CONTROL_STACK_SIZE-16)) || + (pthread_create + (kid_tid,&attr,(void *(*)(void *))new_thread_trampoline,th))) + r=0; + thread_sigmask(SIG_SETMASK,&oldset,0); + return r; +} + +struct thread *create_thread(lispobj initial_function) { struct thread *th; os_thread_t kid_tid=0; - pthread_attr_t attr; + boolean success; if(linux_no_threads_p) return 0; + th=create_thread_struct(initial_function); if(th==0) return 0; -#ifdef QSHOW_SIGNALS - SHOW("create_thread:waiting on lock"); -#endif - get_spinlock(&thread_start_lock,arch_os_get_current_thread()->os_thread); -#ifdef QSHOW_SIGNALS - SHOW("create_thread:got lock"); -#endif - /* The new thread inherits the restrictive signal mask set here, - * and enables signals again when it is set up properly. */ + + /* 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 + 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); - if((pthread_attr_init(&attr)) || - (pthread_attr_setstack(&attr,th->control_stack_start, - THREAD_CONTROL_STACK_SIZE-16)) || - (pthread_create - (&kid_tid,&attr,(void *(*)(void *))new_thread_trampoline,th))) - kid_tid=0; + gc_alloc_update_page_tables(0, &th->alloc_region); thread_sigmask(SIG_SETMASK,&oldset,0); } - if(kid_tid>0) { - link_thread(th,kid_tid); - /* it's started and initialized, it's safe to gc */ - release_spinlock(&thread_start_lock); -#ifdef QSHOW_SIGNALS - SHOW("create_thread:released lock"); -#endif - /* by now the kid might have already exited */ - return kid_tid; - } else { - release_spinlock(&thread_start_lock); -#ifdef QSHOW_SIGNALS - SHOW("create_thread:released lock(failure)"); -#endif - os_invalidate((os_vm_address_t) th->control_stack_start, - ((sizeof (lispobj)) - * (th->control_stack_end-th->control_stack_start)) + - BINDING_STACK_SIZE+ALIEN_STACK_SIZE+dynamic_values_bytes+ - 32*SIGSTKSZ); - return 0; - } -} #endif - -struct thread *find_thread_by_os_thread(os_thread_t tid) -{ - struct thread *th; - for_each_thread(th) - if(th->os_thread==tid) return th; - return 0; + 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); } -#if defined LISP_FEATURE_SB_THREAD -/* This is not needed unless #+SB-THREAD, as there's a trivial null - * unithread definition. */ - -void reap_dead_threads() +/* 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) { - struct thread *th,*next,*prev=0; - th=all_threads; - while(th) { - next=th->next; - if(th->state==STATE_DEAD) { -#ifdef LISP_FEATURE_GENCGC - gc_alloc_update_page_tables(0, &th->alloc_region); -#endif - get_spinlock(&all_threads_lock,th->os_thread); - if(prev) prev->next=next; - else all_threads=next; - release_spinlock(&all_threads_lock); - if(th->tls_cookie>=0) arch_os_thread_cleanup(th); - os_invalidate((os_vm_address_t) th->control_stack_start, - ((sizeof (lispobj)) - * (th->control_stack_end-th->control_stack_start)) + - BINDING_STACK_SIZE+ALIEN_STACK_SIZE+dynamic_values_bytes+ - 32*SIGSTKSZ); - } else - prev=th; - th=next; + int r; + /* The man page does not mention EAGAIN as a valid return value + * for either pthread_kill or kill. But that's theory, this is + * practice. By waiting here we assume that the delivery of this + * signal is not necessary for the delivery of the signals in the + * queue. In other words, we _assume_ there are no deadlocks. */ + while ((r=pthread_kill(os_thread,signo))==EAGAIN) { + /* wait a bit then try again in the hope of the rt signal + * queue not being full */ + FSHOW_SIGNAL((stderr,"/rt signal queue full\n")); + /* FIXME: some kind of backoff (random, exponential) would be + * nice. */ + sleep(1); } + return r; } -int interrupt_thread(os_thread_t tid, lispobj function) +int interrupt_thread(struct thread *th, lispobj function) { - struct thread *th; - for_each_thread(th) - if((th->os_thread==tid) && (th->state != STATE_DEAD)) { - /* 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. - */ - struct cons *c; - int kill_status; - /* mask the signals in case this thread is being interrupted */ - sigset_t newset,oldset; - sigemptyset(&newset); - sigaddset_blockable(&newset); - thread_sigmask(SIG_BLOCK, &newset, &oldset); - - get_spinlock(&th->interrupt_fun_lock, - (int)arch_os_get_current_thread()); - kill_status=thread_kill(th->os_thread,SIG_INTERRUPT_THREAD); - if(kill_status==0) { - c=alloc_cons(function,th->interrupt_fun); - th->interrupt_fun=c; - } - release_spinlock(&th->interrupt_fun_lock); - thread_sigmask(SIG_SETMASK,&oldset,0); - return (kill_status ? -1 : 0); - } - errno=EPERM; return -1; + /* 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)); + } + } } -/* stopping the world is a two-stage process. From this thread we signal +/* stopping the world is a two-stage process. From this thread we signal * all the others with SIG_STOP_FOR_GC. The handler for this signal does - * the usual pseudo-atomic checks (we don't want to stop a thread while + * the usual pseudo-atomic checks (we don't want to stop a thread while * it's in the middle of allocation) then waits for another SIG_STOP_FOR_GC. */ +/* To avoid deadlocks when gc stops the world all clients of each + * mutex must enable or disable SIG_STOP_FOR_GC for the duration of + * holding the lock, but they must agree on which. */ void gc_stop_the_world() { struct thread *p,*th=arch_os_get_current_thread(); -#ifdef QSHOW_SIGNALS - SHOW("gc_stop_the_world:begin"); -#endif + int status; + 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. */ - get_spinlock(&thread_start_lock,th->os_thread); -#ifdef QSHOW_SIGNALS - SHOW("gc_stop_the_world:locked"); -#endif + pthread_mutex_lock(&all_threads_lock); \ + 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->os_thread!=0) && (p->state==STATE_RUNNING)) { - p->state=STATE_STOPPING; - if(thread_kill(p->os_thread,SIG_STOP_FOR_GC)==-1) { - /* FIXME: we can't kill the thread; assume because it died - * already */ - p->state=STATE_DEAD; + if((p!=th) && (p->state==STATE_RUNNING)) { + FSHOW_SIGNAL((stderr,"/gc_stop_the_world: suspending %lu\n", + 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", + p->os_thread,status,strerror(status)); } } } -#ifdef QSHOW_SIGNALS - SHOW("gc_stop_the_world:signals sent"); -#endif - /* wait for the running threads to stop */ + FSHOW_SIGNAL((stderr,"/gc_stop_the_world:signals sent\n")); + /* wait for the running threads to stop or finish */ for(p=all_threads;p;) { - if((p==th) || (p->os_thread==0) || (p->state==STATE_STARTING) || - (p->state==STATE_DEAD) || (p->state==STATE_STOPPED)) { + 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 { + sched_yield(); } } -#ifdef QSHOW_SIGNALS - SHOW("gc_stop_the_world:end"); -#endif + FSHOW_SIGNAL((stderr,"/gc_stop_the_world:end\n")); } void gc_start_the_world() { struct thread *p,*th=arch_os_get_current_thread(); + int status; /* 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 - * restarting; there can be threads just starting from before - * gc_stop_the_world, though */ -#ifdef QSHOW_SIGNALS - SHOW("gc_start_the_world:begin"); -#endif + * restarting */ + FSHOW_SIGNAL((stderr,"/gc_start_the_world:begin\n")); for(p=all_threads;p;p=p->next) { - if((p!=th) && (p->os_thread!=0) && (p->state!=STATE_STARTING) && - (p->state!=STATE_DEAD)) { - if(p->state!=STATE_STOPPED) { - lose("gc_start_the_world: wrong thread state is %ld\n", + gc_assert(p->os_thread!=0); + if((p!=th) && (p->state!=STATE_DEAD)) { + if(p->state!=STATE_SUSPENDED) { + lose("gc_start_the_world: wrong thread state is %d\n", fixnum_value(p->state)); } - thread_kill(p->os_thread,SIG_STOP_FOR_GC); - } - } - /* we must wait for all threads to leave stopped state else we - * risk signal accumulation and lose any meaning of - * thread->state */ - for(p=all_threads;p;) { - gc_assert(p->state!=STATE_STOPPING); - if((p==th) || (p->os_thread==0) || (p->state!=STATE_STOPPED)) { - p=p->next; + FSHOW_SIGNAL((stderr, "/gc_start_the_world: resuming %lu\n", + p->os_thread)); + p->state=STATE_RUNNING; + status=kill_thread_safely(p->os_thread,SIG_STOP_FOR_GC); + if (status) { + lose("cannot resume thread=%lu: %d, %s", + p->os_thread,status,strerror(status)); + } } } - release_spinlock(&thread_start_lock); -#ifdef QSHOW_SIGNALS - SHOW("gc_start_the_world:end"); -#endif + /* If we waited here until all threads leave STATE_SUSPENDED, then + * 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); \ + FSHOW_SIGNAL((stderr,"/gc_start_the_world:end\n")); } #endif