X-Git-Url: http://repo.macrolet.net/gitweb/?a=blobdiff_plain;f=src%2Fruntime%2Fthread.c;h=adeada39f8b908aa7a2f3e28da4d0480ee8ab23a;hb=92c8db80e039f60623e53a0b9355cf0a9ec49f3d;hp=206511aa54a5a82d82286b68857b378b4e20358f;hpb=44dbdff97fbcc1e5b12e1330da92c5a3dcb94a3b;p=sbcl.git diff --git a/src/runtime/thread.c b/src/runtime/thread.c index 206511a..adeada3 100644 --- a/src/runtime/thread.c +++ b/src/runtime/thread.c @@ -1,156 +1,204 @@ #include #include +#include #include #include #include #include -#ifndef CLONE_PARENT /* lameass glibc 2.2 doesn't define this */ -#define CLONE_PARENT 0x00008000 /* even though the manpage documents it */ -#endif -#include "runtime.h" +#include +#include + #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" #include "os.h" #include "globals.h" -#ifdef LISP_FEATURE_GENCGC -#include "gencgc.h" -#endif #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 */ +int dynamic_values_bytes=4096*sizeof(lispobj); /* same for all threads */ struct thread *all_threads; volatile lispobj all_threads_lock; -volatile int countdown_to_gc; extern struct interrupt_data * global_interrupt_data; +extern int linux_no_threads_p; -void get_spinlock(lispobj *word,int value); +#ifdef LISP_FEATURE_SB_THREAD +/* 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(INTERRUPTS_ENABLED,arch_os_get_current_thread()) == NIL) + lose("SIG_STOP_FOR_GC cannot arrive: interrupts disabled\n"); + if (arch_pseudo_atomic_atomic(NULL)) + lose("SIG_STOP_FOR_GC cannot arrive: in pseudo atomic\n"); +} -/* 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 GET_ALL_THREADS_LOCK(name) \ + { \ + sigset_t _newset,_oldset; \ + sigemptyset(&_newset); \ + sigaddset_blockable(&_newset); \ + sigdelset(&_newset,SIG_STOP_FOR_GC); \ + 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)); \ + get_spinlock(&all_threads_lock,(long)arch_os_get_current_thread()); \ + FSHOW_SIGNAL((stderr,"/%s:got lock, thread=%lu\n", \ + name,arch_os_get_current_thread()->os_thread)); -/* set go to 0 to stop the thread before it starts. Convenient if you -* want to attach a debugger to it before it does anything */ -volatile int go=1; +#define RELEASE_ALL_THREADS_LOCK(name) \ + FSHOW_SIGNAL((stderr,"/%s:released lock\n",name)); \ + release_spinlock(&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 int -new_thread_trampoline(struct thread *th) +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; - if(go==0) { - fprintf(stderr, "/pausing 0x%lx(%d,%d) before new_thread_trampoline(0x%lx)\n", - (unsigned long)th,th->pid,getpid(),(unsigned long)function); - while(go==0) ; - fprintf(stderr, "/continue\n"); - } th->unbound_marker = UNBOUND_MARKER_WIDETAG; -#ifdef LISP_FEATURE_SB_THREAD - /* wait here until our thread is linked into all_threads: see below */ - while(th->pid<1) sched_yield(); -#else - if(th->pid < 1) - lose("th->pid not set up right"); -#endif + if(arch_os_thread_init(th)==0) return 1; - if(arch_os_thread_init(th)==0) - return 1; /* failure. no, really */ -#if !defined(LISP_FEATURE_SB_THREAD) && defined(LISP_FEATURE_X86) + 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 return funcall0(function); #endif } +#ifdef LISP_FEATURE_SB_THREAD + +/* 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->os_thread<1) sched_yield(); + + th->state=STATE_RUNNING; + return funcall0(function); +} +#endif /* LISP_FEATURE_SB_THREAD */ + /* 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 */ -pid_t create_thread(lispobj initial_function) { +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; - pid_t kid_pid; /* 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; + THREAD_CONTROL_STACK_SIZE+ + BINDING_STACK_SIZE+ + ALIEN_STACK_SIZE+ + dynamic_values_bytes+ + 32*SIGSTKSZ); + 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); + 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]=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); #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 } + 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_RUNNING; + th->os_thread=0; + th->interrupt_fun=NIL; + th->interrupt_fun_lock=0; + th->state=STATE_STARTING; #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; - /* runtime.c used to set PSEUDO_ATOMIC_ATOMIC =1 globally. I'm not - * sure why, but it appears to help */ - th->pseudo_atomic_atomic=make_fixnum(1); + th->pseudo_atomic_atomic=0; #endif #ifdef LISP_FEATURE_GENCGC gc_set_region_empty(&th->alloc_region); @@ -159,204 +207,279 @@ pid_t create_thread(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); - 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->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; -#ifdef LISP_FEATURE_SB_THREAD -#if defined(LISP_FEATURE_X86) && defined (LISP_FEATURE_LINUX) - kid_pid= - clone(new_thread_trampoline, - (((void*)th->control_stack_start)+THREAD_CONTROL_STACK_SIZE-4), - (((getpid()!=parent_pid)?(CLONE_PARENT):0) - |CLONE_FILES|SIGALRM|CLONE_VM),th); - if(kid_pid<=0) - goto cleanup; -#else -#error this stuff presently only works on x86 Linux -#endif -#else - kid_pid=getpid(); -#endif - get_spinlock(&all_threads_lock,kid_pid); + return th; +} + +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->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 + /* 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->pid,1); - release_spinlock(&all_threads_lock); - th->pid=kid_pid; /* child will not start until this is set */ -#ifndef LISP_FEATURE_SB_THREAD - new_thread_trampoline(all_threads); /* call_into_lisp */ - lose("Clever child? Idiot savant, verging on the."); -#endif - - return th->pid; - 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; + protect_control_stack_guard_page(th,1); + /* child will not start until this is set */ + th->os_thread=kid_tid; } -void destroy_thread (struct thread *th) -{ - /* precondition: the unix task has already been killed and exited. - * This is called by the parent */ -#ifdef LISP_FEATURE_GENCGC - gc_alloc_update_page_tables(0, &th->alloc_region); -#endif - get_spinlock(&all_threads_lock,th->pid); - if(countdown_to_gc>0) countdown_to_gc--; - th->state=STATE_STOPPED; - if(th==all_threads) - all_threads=th->next; - else { - struct thread *th1=all_threads; - while(th1->next!=th) th1=th1->next; - 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); +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"); } +#ifdef LISP_FEATURE_SB_THREAD -struct thread *find_thread_by_pid(pid_t pid) -{ - struct thread *th; - for_each_thread(th) - if(th->pid==pid) return th; - return 0; -} +#ifndef __USE_XOPEN2K +extern int pthread_attr_setstack (pthread_attr_t *__attr, void *__stackaddr, + size_t __stacksize); +#endif -/* 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. */ -#if defined LISP_FEATURE_SB_THREAD -void block_sigcont(void) +boolean create_os_thread(struct thread *th,os_thread_t *kid_tid) { - /* 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; + /* 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); - sigaddset(&newset,SIG_DEQUEUE); - sigprocmask(SIG_BLOCK, &newset, 0); + 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))) + r=0; + thread_sigmask(SIG_SETMASK,&oldset,0); + return r; } -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); +struct thread *create_thread(lispobj initial_function) { + struct thread *th; + os_thread_t kid_tid=0; + boolean success; + + if(linux_no_threads_p) return 0; + + 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 + 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); + + RELEASE_ALL_THREADS_LOCK("create_thread") + + if (success) + return th; + else + return 0; } -int interrupt_thread(pid_t pid, lispobj function) +/* called from lisp from the thread object finalizer */ +void reap_dead_thread(struct thread *th) { - union sigval sigval; - sigval.sival_int=function; + 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); + release_spinlock(&all_threads_lock); + 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); + 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 sigqueue(pid, SIG_INTERRUPT_THREAD, sigval); +int interrupt_thread(struct thread *th, lispobj function) +{ + /* A thread may also become dead after this test. */ + if ((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. + */ + lispobj c=alloc_cons(function,NIL); + int kill_status; + /* interrupt_thread_handler locks this spinlock with + * interrupts 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. */ + sigset_t newset,oldset; + sigemptyset(&newset); + sigaddset_blockable(&newset); + thread_sigmask(SIG_BLOCK, &newset, &oldset); + get_spinlock(&th->interrupt_fun_lock, + (long)arch_os_get_current_thread()); + kill_status=thread_kill(th->os_thread,SIG_INTERRUPT_THREAD); + if(kill_status==0) { + ((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); + return (kill_status ? -1 : 0); + } + errno=EPERM; return -1; } -/* 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 thread 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. - * At any given time, countdown_to_gc should reflect the number of threads - * signalled but which haven't yet come to rest +/* 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 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(); - struct thread *tail=0; - int finished=0; - do { - get_spinlock(&all_threads_lock,th->pid); - if(tail!=all_threads) { - /* new threads always get consed onto the front of all_threads, - * and may be created by any thread that we haven't signalled - * yet or hasn't received our signal and stopped yet. So, check - * for them on each time around */ - for(p=all_threads;p!=tail;p=p->next) { - if(p==th) continue; - /* if the head of all_threads is removed during - * gc_stop_the_world, we may take a second trip through the - * list and end up counting twice as many threads to wait for - * as actually exist */ - if(p->state!=STATE_RUNNING) continue; - countdown_to_gc++; - p->state=STATE_STOPPING; - /* Note no return value check from kill(). If the - * thread had been reaped already, we kill it and - * increment countdown_to_gc anyway. This is to avoid - * complicating the logic in destroy_thread, which would - * otherwise have to know whether the thread died before or - * after it was killed - */ - kill(p->pid,SIG_STOP_FOR_GC); - } - tail=all_threads; - } else { - finished=(countdown_to_gc==0); - } - release_spinlock(&all_threads_lock); - sched_yield(); - } while(!finished); + 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(&all_threads_lock,(long)th); + 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)); + if(thread_kill(p->os_thread,SIG_STOP_FOR_GC)==-1) { + /* we can't kill the thread; assume because it died + * since we last checked */ + p->state=STATE_DEAD; + FSHOW_SIGNAL((stderr,"/gc_stop_the_world:assuming %lu dead\n", + p->os_thread)); + } + } + } + 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; + } + } + FSHOW_SIGNAL((stderr,"/gc_stop_the_world:end\n")); } void gc_start_the_world() { struct thread *p,*th=arch_os_get_current_thread(); - get_spinlock(&all_threads_lock,th->pid); + /* 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 */ + FSHOW_SIGNAL((stderr,"/gc_start_the_world:begin\n")); for(p=all_threads;p;p=p->next) { - if(p==th) continue; - p->state=STATE_RUNNING; - kill(p->pid,SIGCONT); + 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)); + thread_kill(p->os_thread,SIG_STOP_FOR_GC); + } + } + /* we must wait for all threads to leave suspended state else we + * risk signal accumulation and lose any meaning of + * thread->state */ + for(p=all_threads;p;) { + if((p==th) || (p->state!=STATE_SUSPENDED)) { + p=p->next; + } } release_spinlock(&all_threads_lock); + FSHOW_SIGNAL((stderr,"/gc_start_the_world:end\n")); } #endif