X-Git-Url: http://repo.macrolet.net/gitweb/?a=blobdiff_plain;f=src%2Fruntime%2Fthread.c;h=98a897cf3a394d2ac31c0562ef61633aa04283ba;hb=d4d6c4b16a3655ce99a87d43f411391363531260;hp=c9558a5753309520d1ec6b63003be07dad399094;hpb=a4f4342197637938de6b1a835281e7c40136eed5;p=sbcl.git diff --git a/src/runtime/thread.c b/src/runtime/thread.c index c9558a5..98a897c 100644 --- a/src/runtime/thread.c +++ b/src/runtime/thread.c @@ -8,9 +8,10 @@ #include #include -#include "runtime.h" #include "sbcl.h" -#include "validate.h" /* for CONTROL_STACK_SIZE etc */ +#include "runtime.h" +#include "validate.h" /* for CONTROL_STACK_SIZE etc */ +#include "alloc.h" #include "thread.h" #include "arch.h" #include "target-arch-os.h" @@ -19,130 +20,238 @@ #include "dynbind.h" #include "genesis/cons.h" #include "genesis/fdefn.h" +#include "interr.h" /* for lose() */ +#include "gc-internal.h" + #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; +struct freeable_stack { + os_thread_t os_thread; + os_vm_address_t stack; +}; + +static struct freeable_stack * volatile freeable_stack = 0; + +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; +#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) { lispobj function; +#if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64) lispobj *args = NULL; - function = th->unbound_marker; - th->unbound_marker = UNBOUND_MARKER_WIDETAG; +#endif + 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(); + protect_control_stack_guard_page(1); - if(th->pid < 1) lose("th->pid not set up right"); - th->state=STATE_RUNNING; -#if defined(LISP_FEATURE_X86) +#if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64) return call_into_lisp_first_time(function,args,0); #else return funcall0(function); #endif } -/* this is the first thing that clone() 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 bookkeeping needs to be done - */ +#define THREAD_STRUCT_SIZE (THREAD_CONTROL_STACK_SIZE + BINDING_STACK_SIZE + \ + ALIEN_STACK_SIZE + dynamic_values_bytes + \ + 32 * SIGSTKSZ) #ifdef LISP_FEATURE_SB_THREAD + +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 %lu\n", 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)); + } +} + +/* 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 + * bookkeeping needs to be done + */ int new_thread_trampoline(struct thread *th) { lispobj function; - function = th->unbound_marker; - th->unbound_marker = UNBOUND_MARKER_WIDETAG; - if(arch_os_thread_init(th)==0) return 1; - - /* wait here until our thread is linked into all_threads: see below */ - while(th->pid<1) sched_yield(); + int result; + 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) { + /* FIXME: handle error */ + lose("arch_os_thread_init failed\n"); + } - th->state=STATE_RUNNING; - return funcall0(function); + 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). */ + pthread_mutex_lock(&all_threads_lock); + link_thread(th); + pthread_mutex_unlock(&all_threads_lock); + + result = funcall0(function); + 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. */ + pthread_mutex_lock(&all_threads_lock); + gc_alloc_update_page_tables(0, &th->alloc_region); + unlink_thread(th); + pthread_mutex_unlock(&all_threads_lock); + + if(th->tls_cookie>=0) arch_os_thread_cleanup(th); + os_invalidate((os_vm_address_t)th->interrupt_data, + (sizeof (struct interrupt_data))); + free_thread_stack_later(th); + FSHOW((stderr,"/exiting thread %lu\n", thread_self())); + return result; } + #endif /* LISP_FEATURE_SB_THREAD */ +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; +#ifdef LISP_FEATURE_SB_THREAD + int i; +#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 */ - spaces=os_validate(0, - THREAD_CONTROL_STACK_SIZE+ - BINDING_STACK_SIZE+ - ALIEN_STACK_SIZE+ - dynamic_values_bytes+ - 32*SIGSTKSZ - ); - if(!spaces) goto cleanup; + spaces=os_validate(0, THREAD_STRUCT_SIZE); + if(!spaces) + 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); - th=&per_thread->thread; - 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]=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); + 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)) - - 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 } +#endif + 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->pid=0; - th->state=STATE_STOPPED; + th->os_thread=0; + th->state=STATE_RUNNING; #ifdef LISP_FEATURE_STACK_GROWS_DOWNWARD_NOT_UPWARD th->alien_stack_pointer=((void *)th->alien_stack_start - + ALIEN_STACK_SIZE-4); /* naked 4. FIXME */ + + ALIEN_STACK_SIZE-N_WORD_BYTES); #else th->alien_stack_pointer=((void *)th->alien_stack_start); #endif -#ifdef LISP_FEATURE_X86 +#if defined(LISP_FEATURE_X86) || defined (LISP_FEATURE_X86_64) th->pseudo_atomic_interrupted=0; th->pseudo_atomic_atomic=0; #endif @@ -153,248 +262,215 @@ 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 */ - SetSymbolValue(BINDING_STACK_START,th->binding_stack_start,th); - SetSymbolValue(CONTROL_STACK_START,th->control_stack_start,th); - SetSymbolValue(CONTROL_STACK_END,th->control_stack_end,th); -#ifdef LISP_FEATURE_X86 - SetSymbolValue(BINDING_STACK_POINTER,th->binding_stack_pointer,th); - SetSymbolValue(ALIEN_STACK,th->alien_stack_pointer,th); - SetSymbolValue(PSEUDO_ATOMIC_ATOMIC,th->pseudo_atomic_atomic,th); + SetSymbolValue(BINDING_STACK_START,(lispobj)th->binding_stack_start,th); + SetSymbolValue(CONTROL_STACK_START,(lispobj)th->control_stack_start,th); + SetSymbolValue(CONTROL_STACK_END,(lispobj)th->control_stack_end,th); +#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); #else 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=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; + th->interrupt_data = (struct interrupt_data *) + os_validate(0,(sizeof (struct interrupt_data))); + 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; - cleanup: - /* if(th && th->tls_cookie>=0) os_free_tls_pointer(th); */ - if(spaces) os_invalidate(spaces, - THREAD_CONTROL_STACK_SIZE+BINDING_STACK_SIZE+ - ALIEN_STACK_SIZE+dynamic_values_bytes); - return 0; -} - -void link_thread(struct thread *th,pid_t kid_pid) -{ - sigset_t newset,oldset; - sigemptyset(&newset); - sigaddset_blockable(&newset); - sigprocmask(SIG_BLOCK, &newset, &oldset); - - get_spinlock(&all_threads_lock,kid_pid); - th->next=all_threads; - all_threads=th; - /* note that th->pid is 0 at this time. We rely on all_threads_lock - * to ensure that we don't have >1 thread with pid=0 on the list at once - */ - protect_control_stack_guard_page(th->pid,1); - release_spinlock(&all_threads_lock); - - sigprocmask(SIG_SETMASK,&oldset,0); - th->pid=kid_pid; /* child will not start until this is set */ } void create_initial_thread(lispobj initial_function) { struct thread *th=create_thread_struct(initial_function); - pid_t kid_pid=getpid(); - if(th && kid_pid>0) { - link_thread(th,kid_pid); - initial_thread_trampoline(all_threads); /* no return */ + if(th) { + initial_thread_trampoline(th); /* no return */ } else lose("can't create initial thread"); } #ifdef LISP_FEATURE_SB_THREAD -pid_t create_thread(lispobj initial_function) { - struct thread *th=create_thread_struct(initial_function); - pid_t kid_pid=0; - if(th==0) return 0; - kid_pid=clone(new_thread_trampoline, - (((void*)th->control_stack_start)+ - THREAD_CONTROL_STACK_SIZE-4), - CLONE_FILES|SIG_THREAD_EXIT|CLONE_VM,th); - - if(kid_pid>0) { - link_thread(th,kid_pid); - return th->pid; - } else { - 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; - } -} +#ifndef __USE_XOPEN2K +extern int pthread_attr_setstack (pthread_attr_t *__attr, void *__stackaddr, + size_t __stacksize); #endif -/* unused */ -void destroy_thread (struct thread *th) +boolean create_os_thread(struct thread *th,os_thread_t *kid_tid) { - /* precondition: the unix task has already been killed and exited. - * This is called by the parent or some other thread */ -#ifdef LISP_FEATURE_GENCGC - gc_alloc_update_page_tables(0, &th->alloc_region); -#endif - get_spinlock(&all_threads_lock,th->pid); - th->unbound_marker=0; /* for debugging */ - if(th==all_threads) - all_threads=th->next; - else { - struct thread *th1=all_threads; - while(th1 && th1->next!=th) th1=th1->next; - if(th1) th1->next=th->next; /* unlink */ - } - release_spinlock(&all_threads_lock); - if(th && 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); + /* 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, 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)) || + (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 *find_thread_by_pid(pid_t pid) -{ +os_thread_t create_thread(lispobj initial_function) { struct thread *th; - for_each_thread(th) - if(th->pid==pid) return th; - return 0; -} + os_thread_t kid_tid; -#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() -{ - struct thread *th,*next,*prev=0; - th=all_threads; - while(th) { - next=th->next; - if(th->state==STATE_DEAD) { - funcall1(SymbolFunction(HANDLE_THREAD_EXIT),make_fixnum(th->pid)); -#ifdef LISP_FEATURE_GENCGC - gc_alloc_update_page_tables(0, &th->alloc_region); -#endif - get_spinlock(&all_threads_lock,th->pid); - 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; - } -} + if(linux_no_threads_p) return 0; -/* These are not needed unless #+SB-THREAD, and since sigwaitinfo() - * doesn't seem to be easily available everywhere (OpenBSD...) it's - * more trouble than it's worth to compile it when not needed. */ -void block_sigcont(void) -{ - /* don't allow ourselves to receive SIGCONT while we're in the - * "ambiguous" state of being on the queue but not actually stopped. - */ - sigset_t newset; - sigemptyset(&newset); - sigaddset(&newset,SIG_DEQUEUE); - sigprocmask(SIG_BLOCK, &newset, 0); -} + /* 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; -void unblock_sigcont_and_sleep(void) -{ - sigset_t set; - sigemptyset(&set); - sigaddset(&set,SIG_DEQUEUE); - do { - errno=0; - sigwaitinfo(&set,0); - }while(errno==EINTR); - sigprocmask(SIG_UNBLOCK,&set,0); + if (create_os_thread(th,&kid_tid)) { + return kid_tid; + } else { + free_thread_struct(th); + return 0; + } } -int interrupt_thread(pid_t pid, lispobj function) +/* 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) { - union sigval sigval; - sigval.sival_int=function; - - return sigqueue(pid, SIG_INTERRUPT_THREAD, sigval); + 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 signal_thread_to_dequeue (pid_t pid) +int signal_interrupt_thread(os_thread_t os_thread) { - return kill (pid, SIG_DEQUEUE); + 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", + 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 - * it's in the middle of allocation) then kills _itself_ with SIGSTOP. + * 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() { - /* stop all other threads by sending them SIG_STOP_FOR_GC */ struct thread *p,*th=arch_os_get_current_thread(); - pid_t old_pid; - int finished; - do { - finished=1; - for(p=all_threads,old_pid=p->pid; p; p=p->next) { - if(p==th) continue; - if(p->state==STATE_RUNNING) { - p->state=STATE_STOPPING; - kill(p->pid,SIG_STOP_FOR_GC); - } - if((p->state!=STATE_STOPPED) && - (p->state!=STATE_DEAD)) { - finished=0; - } - } - if(old_pid!=all_threads->pid) { - finished=0; - } - } while(!finished); + 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. */ + 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) { + gc_assert(p->os_thread != 0); + 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)); + } + } + } + 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->state==STATE_RUNNING)) { + sched_yield(); + } else { + p=p->next; + } + } + 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_lock, but it won't have been stopped so won't need + * 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 */ + FSHOW_SIGNAL((stderr,"/gc_start_the_world:begin\n")); for(p=all_threads;p;p=p->next) { - if((p==th) || (p->state==STATE_DEAD)) continue; - p->state=STATE_RUNNING; - kill(p->pid,SIG_STOP_FOR_GC); + 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)); + } + 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)); + } + } } + /* 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