X-Git-Url: http://repo.macrolet.net/gitweb/?a=blobdiff_plain;f=src%2Fruntime%2Fthread.c;h=b10ea3c022fa84d7f5bff0d26339ed40cfb52d42;hb=ebc0f0ebf9efd39519ab86ba28c33abdb25443e0;hp=3614c80b014777e7a4aaf8a8eebdd35525de7c6a;hpb=0b5119848b6b8713e473fa669356645747e11dbd;p=sbcl.git diff --git a/src/runtime/thread.c b/src/runtime/thread.c index 3614c80..b10ea3c 100644 --- a/src/runtime/thread.c +++ b/src/runtime/thread.c @@ -1,14 +1,30 @@ +/* + * 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 -#include "sbcl.h" #include "runtime.h" #include "validate.h" /* for CONTROL_STACK_SIZE etc */ #include "alloc.h" @@ -23,39 +39,56 @@ #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 + #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; +#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 -#endif - static void link_thread(struct thread *th) { - th->os_thread=thread_self(); if (all_threads) all_threads->prev=th; th->next=all_threads; th->prev=0; all_threads=th; - protect_control_stack_guard_page(1); } #ifdef LISP_FEATURE_SB_THREAD @@ -82,6 +115,10 @@ initial_thread_trampoline(struct thread *th) 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 defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64) return call_into_lisp_first_time(function,args,0); @@ -96,6 +133,60 @@ initial_thread_trampoline(struct thread *th) #ifdef LISP_FEATURE_SB_THREAD +#ifdef QUEUE_FREEABLE_THREAD_STACKS + +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); + } +} + +#else static void free_thread_stack_later(struct thread *thread_to_be_cleaned_up) { @@ -111,7 +202,7 @@ free_thread_stack_later(struct thread *thread_to_be_cleaned_up) swap_lispobjs((lispobj *)(void *)&freeable_stack, (lispobj)new_freeable_stack); if (new_freeable_stack) { - FSHOW((stderr,"/reaping %lu\n", new_freeable_stack->os_thread)); + 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. */ @@ -121,6 +212,7 @@ free_thread_stack_later(struct thread *thread_to_be_cleaned_up) 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 @@ -131,7 +223,7 @@ 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; @@ -140,28 +232,42 @@ new_thread_trampoline(struct thread *th) lose("arch_os_thread_init failed\n"); } + 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. */ - pthread_mutex_lock(&all_threads_lock); + * 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); - pthread_mutex_unlock(&all_threads_lock); + lock_ret = pthread_mutex_unlock(&all_threads_lock); + gc_assert(lock_ret == 0); 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); + 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); - gc_alloc_update_page_tables(0, &th->alloc_region); if(th->tls_cookie>=0) arch_os_thread_cleanup(th); os_invalidate((os_vm_address_t)th->interrupt_data, (sizeof (struct interrupt_data))); + +#ifdef QUEUE_FREEABLE_THREAD_STACKS + queue_freeable_thread_stack(th); +#else free_thread_stack_later(th); - FSHOW((stderr,"/exiting thread %lu\n", thread_self())); +#endif + + FSHOW((stderr,"/exiting thread %p\n", thread_self())); return result; } @@ -187,6 +293,9 @@ 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 /* 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 @@ -200,17 +309,14 @@ 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; + 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); @@ -230,8 +336,8 @@ create_thread_struct(lispobj initial_function) { 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; @@ -303,7 +409,7 @@ void create_initial_thread(lispobj initial_function) { struct thread *th=create_thread_struct(initial_function); if(th) { initial_thread_trampoline(th); /* no return */ - } else lose("can't create initial thread"); + } else lose("can't create initial thread\n"); } #ifdef LISP_FEATURE_SB_THREAD @@ -320,6 +426,10 @@ 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")); + sigemptyset(&newset); /* Blocking deferrable signals is enough, no need to block * SIG_STOP_FOR_GC because the child process is not linked onto @@ -327,12 +437,31 @@ boolean create_os_thread(struct thread *th,os_thread_t *kid_tid) 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); return r; } @@ -341,9 +470,11 @@ os_thread_t create_thread(lispobj initial_function) { struct thread *th; os_thread_t kid_tid; - if(linux_no_threads_p) return 0; - - 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; if (create_os_thread(th,&kid_tid)) { @@ -383,7 +514,7 @@ int signal_interrupt_thread(os_thread_t os_thread) } else if (status == ESRCH) { return -1; } else { - lose("cannot send SIG_INTERRUPT_THREAD to thread=%lu: %d, %s", + lose("cannot send SIG_INTERRUPT_THREAD to thread=%lu: %d, %s\n", os_thread, status, strerror(status)); } } @@ -400,24 +531,28 @@ int signal_interrupt_thread(os_thread_t os_thread) void gc_stop_the_world() { struct thread *p,*th=arch_os_get_current_thread(); - int status; + int status, lock_ret; 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) { + 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 %lu\n", - p->os_thread)); + 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)); } } @@ -425,12 +560,11 @@ 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); - 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")); @@ -439,7 +573,7 @@ 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 @@ -455,9 +589,14 @@ 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)); } } @@ -466,7 +605,10 @@ 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); + FSHOW_SIGNAL((stderr,"/gc_start_the_world:end\n")); } #endif