X-Git-Url: http://repo.macrolet.net/gitweb/?a=blobdiff_plain;f=src%2Fruntime%2Fthread.c;h=ac1ed7932166adda340ebf7dd663cdefba655285;hb=137ba2db2d362f03754ccd080ddbe96f7e3c5dc7;hp=2c3367c3f7710d54eea6055083591fb51e5f7934;hpb=8e9908b6f79b7bb52a281cfcbf0712d4b914f3cf;p=sbcl.git diff --git a/src/runtime/thread.c b/src/runtime/thread.c index 2c3367c..ac1ed79 100644 --- a/src/runtime/thread.c +++ b/src/runtime/thread.c @@ -32,8 +32,7 @@ #endif #include "runtime.h" -#include "validate.h" /* for CONTROL_STACK_SIZE etc */ -#include "alloc.h" +#include "validate.h" /* for BINDING_STACK_SIZE etc */ #include "thread.h" #include "arch.h" #include "target-arch-os.h" @@ -43,6 +42,7 @@ #include "genesis/cons.h" #include "genesis/fdefn.h" #include "interr.h" /* for lose() */ +#include "alloc.h" #include "gc-internal.h" #ifdef LISP_FEATURE_WIN32 @@ -54,7 +54,8 @@ #endif #if defined(LISP_FEATURE_DARWIN) && defined(LISP_FEATURE_SB_THREAD) -#define QUEUE_FREEABLE_THREAD_STACKS +#define DELAY_THREAD_POST_MORTEM 5 +#define LOCK_CREATE_THREAD #endif #ifdef LISP_FEATURE_FREEBSD @@ -62,27 +63,25 @@ #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; +#ifdef LISP_FEATURE_SB_THREAD +struct thread_post_mortem { +#ifdef DELAY_THREAD_POST_MORTEM + struct thread_post_mortem *next; #endif os_thread_t os_thread; - os_vm_address_t stack; + pthread_attr_t *os_attr; + os_vm_address_t os_address; }; - -#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; +#ifdef DELAY_THREAD_POST_MORTEM +static int pending_thread_post_mortem_count = 0; +pthread_mutex_t thread_post_mortem_lock = PTHREAD_MUTEX_INITIALIZER; +#endif +static struct thread_post_mortem * volatile pending_thread_post_mortem = 0; #endif -int dynamic_values_bytes=4096*sizeof(lispobj); /* same for all threads */ -struct thread * volatile all_threads; +int dynamic_values_bytes=TLS_SIZE*sizeof(lispobj); /* same for all threads */ +struct thread *all_threads; extern struct interrupt_data * global_interrupt_data; #ifdef LISP_FEATURE_SB_THREAD @@ -90,6 +89,10 @@ pthread_mutex_t all_threads_lock = PTHREAD_MUTEX_INITIALIZER; #ifdef LOCK_CREATE_THREAD static pthread_mutex_t create_thread_lock = PTHREAD_MUTEX_INITIALIZER; #endif +#ifdef LISP_FEATURE_GCC_TLS +__thread struct thread *current_thread; +#endif +pthread_key_t lisp_thread = 0; #endif #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64) @@ -125,13 +128,25 @@ initial_thread_trampoline(struct thread *th) #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64) lispobj *args = NULL; #endif +#ifdef LISP_FEATURE_SB_THREAD + pthread_setspecific(lisp_thread, (void *)1); +#endif +#if defined(LISP_FEATURE_SB_THREAD) && defined(LISP_FEATURE_PPC) + /* SIG_STOP_FOR_GC defaults to blocked on PPC? */ + unblock_gc_signals(0,0); +#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(); #ifndef LISP_FEATURE_WIN32 - protect_control_stack_guard_page(1); + protect_control_stack_hard_guard_page(1, NULL); + protect_binding_stack_hard_guard_page(1, NULL); + protect_alien_stack_hard_guard_page(1, NULL); + protect_control_stack_guard_page(1, NULL); + protect_binding_stack_guard_page(1, NULL); + protect_alien_stack_guard_page(1, NULL); #endif #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64) @@ -141,132 +156,106 @@ 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 +#define THREAD_STATE_LOCK_SIZE \ + (sizeof(pthread_mutex_t))+(sizeof(pthread_cond_t)) +#else +#define THREAD_STATE_LOCK_SIZE 0 +#endif -#ifdef QUEUE_FREEABLE_THREAD_STACKS +#define THREAD_STRUCT_SIZE (thread_control_stack_size + BINDING_STACK_SIZE + \ + ALIEN_STACK_SIZE + \ + THREAD_STATE_LOCK_SIZE + \ + dynamic_values_bytes + \ + 32 * SIGSTKSZ + \ + THREAD_ALIGNMENT_BYTES) -static void -queue_freeable_thread_stack(struct thread *thread_to_be_cleaned_up) +#ifdef LISP_FEATURE_SB_THREAD +/* THREAD POST MORTEM CLEANUP + * + * Memory allocated for the thread stacks cannot be reclaimed while + * the thread is still alive, so we need a mechanism for post mortem + * cleanups. FIXME: We actually have three, for historical reasons as + * the saying goes. Do we really need three? Nikodemus guesses that + * not anymore, now that we properly call pthread_attr_destroy before + * freeing the stack. */ + +static struct thread_post_mortem * +plan_thread_post_mortem(struct thread *corpse) { - 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); + if (corpse) { + struct thread_post_mortem *post_mortem = malloc(sizeof(struct thread_post_mortem)); + gc_assert(post_mortem); + post_mortem->os_thread = corpse->os_thread; + post_mortem->os_attr = corpse->os_attr; + post_mortem->os_address = corpse->os_address; +#ifdef DELAY_THREAD_POST_MORTEM + post_mortem->next = NULL; +#endif + return post_mortem; + } else { + /* FIXME: When does this happen? */ + return NULL; } } -#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) +perform_thread_post_mortem(struct thread_post_mortem *post_mortem) { - 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(); +#ifdef CREATE_POST_MORTEM_THREAD + pthread_detach(pthread_self()); +#endif + if (post_mortem) { + gc_assert(!pthread_join(post_mortem->os_thread, NULL)); + gc_assert(!pthread_attr_destroy(post_mortem->os_attr)); + free(post_mortem->os_attr); + os_invalidate(post_mortem->os_address, THREAD_STRUCT_SIZE); + free(post_mortem); } } -#else static void -free_thread_stack_later(struct thread *thread_to_be_cleaned_up) +schedule_thread_post_mortem(struct thread *corpse) { - 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)); + struct thread_post_mortem *post_mortem = NULL; + if (corpse) { + post_mortem = plan_thread_post_mortem(corpse); + +#ifdef DELAY_THREAD_POST_MORTEM + pthread_mutex_lock(&thread_post_mortem_lock); + /* First stick the new post mortem to the end of the queue. */ + if (pending_thread_post_mortem) { + struct thread_post_mortem *next = pending_thread_post_mortem; + while (next->next) { + next = next->next; + } + next->next = post_mortem; + } else { + pending_thread_post_mortem = post_mortem; + } + /* Then, if there are enough things in the queue, clean up one + * from the head -- or increment the count, and null out the + * post_mortem we have. */ + if (pending_thread_post_mortem_count > DELAY_THREAD_POST_MORTEM) { + post_mortem = pending_thread_post_mortem; + pending_thread_post_mortem = post_mortem->next; + } else { + pending_thread_post_mortem_count++; + post_mortem = NULL; + } + pthread_mutex_unlock(&thread_post_mortem_lock); + /* Finally run, the cleanup, if any. */ + perform_thread_post_mortem(post_mortem); +#elif defined(CREATE_POST_MORTEM_THREAD) + gc_assert(!pthread_create(&thread, NULL, perform_thread_post_mortem, post_mortem)); +#else + post_mortem = (struct thread_post_mortem *) + swap_lispobjs((lispobj *)(void *)&pending_thread_post_mortem, + (lispobj)post_mortem); + perform_thread_post_mortem(post_mortem); +#endif } } -#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 @@ -280,6 +269,9 @@ new_thread_trampoline(struct thread *th) int result, lock_ret; FSHOW((stderr,"/creating thread %lu\n", thread_self())); + check_deferrables_blocked_or_lose(0); + check_gc_signals_unblocked_or_lose(0); + pthread_setspecific(lisp_thread, (void *)1); function = th->no_tls_value_marker; th->no_tls_value_marker = NO_TLS_VALUE_MARKER_WIDETAG; if(arch_os_thread_init(th)==0) { @@ -288,11 +280,13 @@ new_thread_trampoline(struct thread *th) } th->os_thread=thread_self(); - protect_control_stack_guard_page(1); + protect_control_stack_guard_page(1, NULL); + protect_binding_stack_guard_page(1, NULL); + protect_alien_stack_guard_page(1, NULL); /* 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). */ + * 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); @@ -302,43 +296,39 @@ new_thread_trampoline(struct thread *th) result = funcall0(function); /* Block GC */ - block_blockable_signals(); - th->state=STATE_DEAD; + block_blockable_signals(0, 0); + set_thread_state(th, 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); + gc_alloc_update_page_tables(BOXED_PAGE_FLAG, &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); + pthread_mutex_destroy(th->state_lock); + pthread_cond_destroy(th->state_cond); + 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(), + mach_port_move_member(current_mach_task, THREAD_STRUCT_TO_EXCEPTION_PORT(th), MACH_PORT_NULL); - mach_port_deallocate(mach_task_self(), + mach_port_deallocate(current_mach_task, THREAD_STRUCT_TO_EXCEPTION_PORT(th)); - mach_port_destroy(mach_task_self(), + mach_port_destroy(current_mach_task, 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())); + schedule_thread_post_mortem(th); + FSHOW((stderr,"/exiting thread %lu\n", thread_self())); return result; } @@ -350,10 +340,16 @@ 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, + os_invalidate((os_vm_address_t) th->os_address, THREAD_STRUCT_SIZE); } +#ifdef LISP_FEATURE_SB_THREAD +/* FIXME: should be MAX_INTERRUPTS -1 ? */ +const unsigned int tls_index_start = + MAX_INTERRUPTS + sizeof(struct thread)/sizeof(lispobj); +#endif + /* 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 @@ -364,48 +360,55 @@ create_thread_struct(lispobj initial_function) { union per_thread_data *per_thread; struct thread *th=0; /* subdue gcc */ void *spaces=0; + void *aligned_spaces=0; #ifdef LISP_FEATURE_SB_THREAD - int i; + unsigned 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 + /* 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 */ + * succeed. SPACES must be appropriately aligned, since the GC + * expects the control stack to start at a page boundary -- and + * the OS may have even more rigorous requirements. We can't rely + * on the alignment passed from os_validate, since that might + * assume the current (e.g. 4k) pagesize, while we calculate with + * the biggest (e.g. 64k) pagesize allowed by the ABI. */ spaces=os_validate(0, THREAD_STRUCT_SIZE); if(!spaces) - return NULL; + return NULL; + /* Aligning up is safe as THREAD_STRUCT_SIZE has + * THREAD_ALIGNMENT_BYTES padding. */ + aligned_spaces = (void *)((((unsigned long)(char *)spaces) + + THREAD_ALIGNMENT_BYTES-1) + &~(unsigned long)(THREAD_ALIGNMENT_BYTES-1)); per_thread=(union per_thread_data *) - (spaces+ - THREAD_CONTROL_STACK_SIZE+ + (aligned_spaces+ + thread_control_stack_size+ BINDING_STACK_SIZE+ - ALIEN_STACK_SIZE); + ALIEN_STACK_SIZE + + THREAD_STATE_LOCK_SIZE); #ifdef LISP_FEATURE_SB_THREAD 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(FREE_TLS_INDEX,tls_index_start << WORD_SHIFT,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)) + (THREAD_SLOT_OFFSET_WORDS(field) << WORD_SHIFT) STATIC_TLS_INIT(BINDING_STACK_START,binding_stack_start); +#ifdef BINDING_STACK_POINTER STATIC_TLS_INIT(BINDING_STACK_POINTER,binding_stack_pointer); +#endif STATIC_TLS_INIT(CONTROL_STACK_START,control_stack_start); STATIC_TLS_INIT(CONTROL_STACK_END,control_stack_end); +#ifdef ALIEN_STACK STATIC_TLS_INIT(ALIEN_STACK,alien_stack_pointer); +#endif #if defined(LISP_FEATURE_X86) || defined (LISP_FEATURE_X86_64) STATIC_TLS_INIT(PSEUDO_ATOMIC_BITS,pseudo_atomic_bits); #endif @@ -414,15 +417,26 @@ create_thread_struct(lispobj initial_function) { #endif th=&per_thread->thread; - th->control_stack_start = spaces; + th->os_address = spaces; + th->control_stack_start = aligned_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->control_stack_guard_page_protected = T; th->alien_stack_start= (lispobj*)((void*)th->binding_stack_start+BINDING_STACK_SIZE); - th->binding_stack_pointer=th->binding_stack_start; + set_binding_stack_pointer(th,th->binding_stack_start); th->this=th; th->os_thread=0; +#ifdef LISP_FEATURE_SB_THREAD + th->os_attr=malloc(sizeof(pthread_attr_t)); + th->state_lock=(pthread_mutex_t *)((void *)th->alien_stack_start + + ALIEN_STACK_SIZE); + pthread_mutex_init(th->state_lock, NULL); + th->state_cond=(pthread_cond_t *)((void *)th->state_lock + + (sizeof(pthread_mutex_t))); + pthread_cond_init(th->state_cond, NULL); +#endif th->state=STATE_RUNNING; #ifdef LISP_FEATURE_STACK_GROWS_DOWNWARD_NOT_UPWARD th->alien_stack_pointer=((void *)th->alien_stack_start @@ -430,12 +444,22 @@ create_thread_struct(lispobj initial_function) { #else th->alien_stack_pointer=((void *)th->alien_stack_start); #endif -#if defined(LISP_FEATURE_X86) || defined (LISP_FEATURE_X86_64) +#if defined(LISP_FEATURE_X86) || defined (LISP_FEATURE_X86_64) || defined(LISP_FEATURE_SB_THREAD) th->pseudo_atomic_bits=0; #endif #ifdef LISP_FEATURE_GENCGC gc_set_region_empty(&th->alloc_region); #endif +#ifdef LISP_FEATURE_SB_THREAD + /* This parallels the same logic in globals.c for the + * single-threaded foreign_function_call_active, KLUDGE and + * all. */ +#if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64) + th->foreign_function_call_active = 0; +#else + th->foreign_function_call_active = 1; +#endif +#endif #ifndef LISP_FEATURE_SB_THREAD /* the tls-points-into-struct-thread trick is only good for threaded @@ -448,12 +472,8 @@ create_thread_struct(lispobj initial_function) { 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_BITS,(lispobj)th->pseudo_atomic_bits,th); -#else - current_binding_stack_pointer=th->binding_stack_pointer; - current_control_stack_pointer=th->control_stack_start; #endif #endif bind_variable(CURRENT_CATCH_BLOCK,make_fixnum(0),th); @@ -461,10 +481,18 @@ create_thread_struct(lispobj initial_function) { bind_variable(FREE_INTERRUPT_CONTEXT_INDEX,make_fixnum(0),th); bind_variable(INTERRUPT_PENDING, NIL,th); bind_variable(INTERRUPTS_ENABLED,T,th); + bind_variable(ALLOW_WITH_INTERRUPTS,T,th); bind_variable(GC_PENDING,NIL,th); + bind_variable(ALLOC_SIGNAL,NIL,th); +#ifdef PINNED_OBJECTS + bind_variable(PINNED_OBJECTS,NIL,th); +#endif #ifdef LISP_FEATURE_SB_THREAD bind_variable(STOP_FOR_GC_PENDING,NIL,th); #endif +#ifndef LISP_FEATURE_C_STACK_IS_CONTROL_STACK + access_control_stack_pointer(th)=th->control_stack_start; +#endif th->interrupt_data = (struct interrupt_data *) os_validate(0,(sizeof (struct interrupt_data))); @@ -473,6 +501,10 @@ create_thread_struct(lispobj initial_function) { return 0; } th->interrupt_data->pending_handler = 0; + th->interrupt_data->gc_blocked_deferrables = 0; +#ifdef LISP_FEATURE_PPC + th->interrupt_data->allocation_trap_context = 0; +#endif th->no_tls_value_marker=initial_function; th->stepping = NIL; @@ -487,10 +519,11 @@ kern_return_t mach_thread_init(mach_port_t thread_exception_port); void create_initial_thread(lispobj initial_function) { struct thread *th=create_thread_struct(initial_function); +#ifdef LISP_FEATURE_SB_THREAD + pthread_key_create(&lisp_thread, 0); +#endif if(th) { #ifdef LISP_FEATURE_MACH_EXCEPTION_HANDLER - kern_return_t ret; - setup_mach_exception_handling_thread(); #endif initial_thread_trampoline(th); /* no return */ @@ -508,112 +541,66 @@ 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; + sigset_t oldset; boolean r=1; - int retcode, initcode, sizecode, addrcode; + int retcode = 0, initcode; 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, 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); + block_deferrable_signals(0, &oldset); -#if defined(LISP_FEATURE_DARWIN) -#define CONTROL_STACK_ADJUST 8192 /* darwin wants page-aligned stacks */ -#else -#define CONTROL_STACK_ADJUST 16 +#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 - 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-CONTROL_STACK_ADJUST)) || -#undef CONTROL_STACK_ADJUST + if((initcode = pthread_attr_init(th->os_attr)) || + /* call_into_lisp_first_time switches the stack for the initial + * thread. For the others, we use this. */ + (pthread_attr_setstack(th->os_attr,th->control_stack_start, + thread_control_stack_size)) || (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)); + (kid_tid,th->os_attr,(void *(*)(void *))new_thread_trampoline,th))) { + FSHOW_SIGNAL((stderr, "init = %d\n", initcode)); + FSHOW_SIGNAL((stderr, "pthread_create returned %d, errno %d\n", + retcode, errno)); 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 + thread_sigmask(SIG_SETMASK,&oldset,0); return r; } os_thread_t create_thread(lispobj initial_function) { - struct thread *th; - os_thread_t kid_tid; + struct thread *th, *thread = arch_os_get_current_thread(); + os_thread_t kid_tid = 0; + + /* Must defend against async unwinds. */ + if (SymbolValue(INTERRUPTS_ENABLED, thread) != NIL) + lose("create_thread is not safe when interrupts are enabled.\n"); /* 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)) { - return kid_tid; - } else { + if (th && !create_os_thread(th,&kid_tid)) { free_thread_struct(th); - return 0; - } -} - -/* Send the signo to os_thread, retry if the rt signal queue is - * full. */ -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 - * 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_interrupt_thread(os_thread_t os_thread) -{ - 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)); + kid_tid = 0; } + return kid_tid; } /* stopping the world is a two-stage process. From this thread we signal @@ -632,32 +619,31 @@ void gc_stop_the_world() #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)); + FSHOW_SIGNAL((stderr,"/gc_stop_the_world:waiting on create_thread_lock\n")); 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)); + FSHOW_SIGNAL((stderr,"/gc_stop_the_world:got create_thread_lock\n")); #endif - FSHOW_SIGNAL((stderr,"/gc_stop_the_world:waiting on lock, thread=%lu\n", - th->os_thread)); + FSHOW_SIGNAL((stderr,"/gc_stop_the_world:waiting on lock\n")); /* keep threads from starting while the world is stopped. */ 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)); + FSHOW_SIGNAL((stderr,"/gc_stop_the_world:got lock\n")); /* 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 %x, os_thread %x\n", - p, p->os_thread)); - status=kill_thread_safely(p->os_thread,SIG_STOP_FOR_GC); + FSHOW_SIGNAL((stderr,"/gc_stop_the_world: thread=%lu, state=%x\n", + p->os_thread, thread_state(p))); + if((p!=th) && ((thread_state(p)==STATE_RUNNING))) { + FSHOW_SIGNAL((stderr,"/gc_stop_the_world: suspending thread %lu\n", + p->os_thread)); + /* We already hold all_thread_lock, P can become DEAD but + * cannot exit, ergo it's safe to use pthread_kill. */ + status=pthread_kill(p->os_thread,SIG_STOP_FOR_GC); if (status==ESRCH) { /* This thread has exited. */ - gc_assert(p->state==STATE_DEAD); + gc_assert(thread_state(p)==STATE_DEAD); } else if (status) { lose("cannot send suspend thread=%lu: %d, %s\n", p->os_thread,status,strerror(status)); @@ -665,13 +651,15 @@ 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;) { - 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; + for(p=all_threads;p;p=p->next) { + if (p!=th) { + FSHOW_SIGNAL + ((stderr, + "/gc_stop_the_world: waiting for thread=%lu: state=%x\n", + p->os_thread, thread_state(p))); + wait_for_thread_state_change(p, STATE_RUNNING); + if (p->state == STATE_RUNNING) + lose("/gc_stop_the_world: unexpected state"); } } FSHOW_SIGNAL((stderr,"/gc_stop_the_world:end\n")); @@ -680,7 +668,7 @@ void gc_stop_the_world() void gc_start_the_world() { struct thread *p,*th=arch_os_get_current_thread(); - int status, lock_ret; + int 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 @@ -688,30 +676,19 @@ void gc_start_the_world() FSHOW_SIGNAL((stderr,"/gc_start_the_world:begin\n")); for(p=all_threads;p;p=p->next) { 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; - -#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\n", - p->os_thread,status,strerror(status)); + if (p!=th) { + lispobj state = thread_state(p); + if (state != STATE_DEAD) { + if(state != STATE_SUSPENDED) { + lose("gc_start_the_world: wrong thread state is %d\n", + fixnum_value(state)); + } + FSHOW_SIGNAL((stderr, "/gc_start_the_world: resuming %lu\n", + p->os_thread)); + set_thread_state(p, STATE_RUNNING); } } } - /* 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. */ lock_ret = pthread_mutex_unlock(&all_threads_lock); gc_assert(lock_ret == 0); @@ -723,3 +700,81 @@ void gc_start_the_world() FSHOW_SIGNAL((stderr,"/gc_start_the_world:end\n")); } #endif + +int +thread_yield() +{ +#ifdef LISP_FEATURE_SB_THREAD + return sched_yield(); +#else + return 0; +#endif +} + +/* If the thread id given does not belong to a running thread (it has + * exited or never even existed) pthread_kill _may_ fail with ESRCH, + * but it is also allowed to just segfault, see + * . + * + * Relying on thread ids can easily backfire since ids are recycled + * (NPTL recycles them extremely fast) so a signal can be sent to + * another process if the one it was sent to exited. + * + * We send signals in two places: signal_interrupt_thread sends a + * signal that's harmless if delivered to another thread, but + * SIG_STOP_FOR_GC is fatal. + * + * For these reasons, we must make sure that the thread is still alive + * when the pthread_kill is called and return if the thread is + * exiting. */ +int +kill_safely(os_thread_t os_thread, int signal) +{ + FSHOW_SIGNAL((stderr,"/kill_safely: %lu, %d\n", os_thread, signal)); + { +#ifdef LISP_FEATURE_SB_THREAD + sigset_t oldset; + struct thread *thread; + /* pthread_kill is not async signal safe and we don't want to be + * interrupted while holding the lock. */ + block_deferrable_signals(0, &oldset); + pthread_mutex_lock(&all_threads_lock); + for (thread = all_threads; thread; thread = thread->next) { + if (thread->os_thread == os_thread) { + int status = pthread_kill(os_thread, signal); + if (status) + lose("kill_safely: pthread_kill failed with %d\n", status); + break; + } + } + pthread_mutex_unlock(&all_threads_lock); + thread_sigmask(SIG_SETMASK,&oldset,0); + if (thread) + return 0; + else + return -1; +#else + int status; + if (os_thread != 0) + lose("kill_safely: who do you want to kill? %d?\n", os_thread); + /* Dubious (as in don't know why it works) workaround for the + * signal sometimes not being generated on darwin. */ +#ifdef LISP_FEATURE_DARWIN + { + sigset_t oldset; + sigprocmask(SIG_BLOCK, &deferrable_sigset, &oldset); + status = raise(signal); + sigprocmask(SIG_SETMASK,&oldset,0); + } +#else + status = raise(signal); +#endif + if (status == 0) { + return 0; + } else { + lose("cannot raise signal %d, %d %s\n", + signal, status, strerror(errno)); + } +#endif + } +}