X-Git-Url: http://repo.macrolet.net/gitweb/?a=blobdiff_plain;f=src%2Fruntime%2Fthread.c;h=84fb9471c856e5b5db663a848dd1a82d9a8234d4;hb=cf507f95509a855a752b6f1771aa06877b8a3b30;hp=1de8a200257e5f6c082154322398daaf4f2f00ce;hpb=629e57661251d3879177bcf52040d1380f7bbfe3;p=sbcl.git diff --git a/src/runtime/thread.c b/src/runtime/thread.c index 1de8a20..84fb947 100644 --- a/src/runtime/thread.c +++ b/src/runtime/thread.c @@ -1,153 +1,433 @@ +/* + * 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 -#ifndef CLONE_PARENT /* lameass glibc 2.2 doesn't define this */ -#define CLONE_PARENT 0x00008000 /* even though the manpage documents it */ +#include +#include +#ifndef LISP_FEATURE_WIN32 +#include #endif + +#ifdef LISP_FEATURE_MACH_EXCEPTION_HANDLER +#include +#include +#include +#endif + #include "runtime.h" -#include "sbcl.h" -#include "validate.h" /* for CONTROL_STACK_SIZE etc */ +#include "validate.h" /* for BINDING_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" -#define ALIEN_STACK_SIZE (1*1024*1024) /* 1Mb size chosen at random */ +#include "genesis/fdefn.h" +#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 DELAY_THREAD_POST_MORTEM 5 +#define LOCK_CREATE_THREAD +#endif + +#ifdef LISP_FEATURE_FREEBSD +#define CREATE_CLEANUP_THREAD +#define LOCK_CREATE_THREAD +#endif + +#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; + pthread_attr_t *os_attr; + os_vm_address_t os_address; +}; -int dynamic_values_bytes=4096*sizeof(lispobj); /* same for all threads */ +#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=TLS_SIZE*sizeof(lispobj); /* same for all threads */ struct thread *all_threads; -lispobj all_threads_lock; extern struct interrupt_data * global_interrupt_data; -void get_spinlock(lispobj *word,int value); +#ifdef LISP_FEATURE_SB_THREAD +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 +#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 - */ +#if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64) +extern lispobj call_into_lisp_first_time(lispobj fun, lispobj *args, int nargs); +#endif -/* 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; +static void +link_thread(struct thread *th) +{ + if (all_threads) all_threads->prev=th; + th->next=all_threads; + th->prev=0; + all_threads=th; +} -int -new_thread_trampoline(struct thread *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; - 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 + 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, NULL); + protect_binding_stack_guard_page(1, NULL); + protect_alien_stack_guard_page(1, NULL); #endif - if(arch_os_thread_init(th)==0) - return 1; /* failure. no, really */ -#if !defined(LISP_FEATURE_SB_THREAD) && 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 } +#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 + +#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) + +#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 (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; + } +} + +static void +perform_thread_post_mortem(struct thread_post_mortem *post_mortem) +{ +#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); + } +} + +static void +schedule_thread_post_mortem(struct thread *corpse) +{ + 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 + } +} + +/* 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; + 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); + 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->os_thread=thread_self(); + 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). */ + lock_ret = pthread_mutex_lock(&all_threads_lock); + gc_assert(lock_ret == 0); + link_thread(th); + lock_ret = pthread_mutex_unlock(&all_threads_lock); + gc_assert(lock_ret == 0); + + result = funcall0(function); + + /* Block GC */ + 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(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(), + THREAD_STRUCT_TO_EXCEPTION_PORT(th), + MACH_PORT_NULL); + mach_port_deallocate(mach_task_self(), + THREAD_STRUCT_TO_EXCEPTION_PORT(th)); + mach_port_destroy(mach_task_self(), + THREAD_STRUCT_TO_EXCEPTION_PORT(th)); +#endif + + schedule_thread_post_mortem(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->os_address, + 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 */ -pid_t create_thread(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; - pid_t kid_pid; + void *aligned_spaces=0; +#ifdef LISP_FEATURE_SB_THREAD + unsigned int i; +#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 */ - spaces=os_validate(0, - THREAD_CONTROL_STACK_SIZE+ - BINDING_STACK_SIZE+ - ALIEN_STACK_SIZE+ - dynamic_values_bytes+ - 32*SIGSTKSZ - ); - if(!spaces) goto cleanup; + * 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; + /* 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+ - BINDING_STACK_SIZE+ - ALIEN_STACK_SIZE); + (aligned_spaces+ + thread_control_stack_size+ + BINDING_STACK_SIZE+ + ALIEN_STACK_SIZE + + THREAD_STATE_LOCK_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_BITS,pseudo_atomic_bits); #endif #undef STATIC_TLS_INIT -#endif } +#endif - th->control_stack_start = spaces; + th=&per_thread->thread; + 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->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->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 - + 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 - 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); +#if defined(LISP_FEATURE_X86) || defined (LISP_FEATURE_X86_64) + th->pseudo_atomic_bits=0; #endif #ifdef LISP_FEATURE_GENCGC gc_set_region_empty(&th->alloc_region); @@ -156,130 +436,309 @@ 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(PSEUDO_ATOMIC_INTERRUPTED,th->pseudo_atomic_interrupted,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_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 +#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(ALLOW_WITH_INTERRUPTS,T,th); + bind_variable(GC_PENDING,NIL,th); + bind_variable(ALLOC_SIGNAL,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->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->interrupt_data->gc_blocked_deferrables = 0; + th->no_tls_value_marker=initial_function; + + th->stepping = NIL; + return th; +} + +#ifdef LISP_FEATURE_MACH_EXCEPTION_HANDLER +mach_port_t setup_mach_exception_handling_thread(); +kern_return_t mach_thread_init(mach_port_t thread_exception_port); - 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(); + +void create_initial_thread(lispobj initial_function) { + struct thread *th=create_thread_struct(initial_function); + if(th) { +#ifdef LISP_FEATURE_MACH_EXCEPTION_HANDLER + setup_mach_exception_handling_thread(); #endif - 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); - all_threads_lock=0; - 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."); + initial_thread_trampoline(th); /* no return */ + } else lose("can't create initial thread\n"); +} + +#ifdef LISP_FEATURE_SB_THREAD + +#ifndef __USE_XOPEN2K +extern int pthread_attr_setstack (pthread_attr_t *__attr, void *__stackaddr, + size_t __stacksize); #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; +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. */ + sigset_t oldset; + boolean r=1; + int retcode = 0, initcode; + + FSHOW_SIGNAL((stderr,"/create_os_thread: creating new thread\n")); + + /* 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. */ + block_deferrable_signals(0, &oldset); + +#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(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,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 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; } -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(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 */ +os_thread_t create_thread(lispobj initial_function) { + 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 && !create_os_thread(th,&kid_tid)) { + free_thread_struct(th); + kid_tid = 0; } - all_threads_lock=0; - if(th && th->tls_cookie>=0) arch_os_thread_cleanup(th); - os_invalidate((os_vm_address_t) th->control_stack_start, - THREAD_CONTROL_STACK_SIZE+BINDING_STACK_SIZE+ - ALIEN_STACK_SIZE+dynamic_values_bytes+ - 32*SIGSTKSZ); + return kid_tid; } +/* 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. + */ -struct thread *find_thread_by_pid(pid_t pid) +/* To avoid deadlocks when gc stops the world all clients of each + * mutex must enable or disable SIG_STOP_FOR_GC for the duration of + * holding the lock, but they must agree on which. */ +void gc_stop_the_world() { - struct thread *th; - for_each_thread(th) - if(th->pid==pid) return th; - return 0; + struct thread *p,*th=arch_os_get_current_thread(); + int status, lock_ret; +#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\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\n")); +#endif + 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\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: 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(thread_state(p)==STATE_DEAD); + } else if (status) { + lose("cannot send suspend thread=%lu: %d, %s\n", + p->os_thread,status,strerror(status)); + } + } + } + FSHOW_SIGNAL((stderr,"/gc_stop_the_world:signals sent\n")); + 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")); } -void block_sigcont(void) +void gc_start_the_world() { - /* 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,SIGCONT); - sigprocmask(SIG_BLOCK, &newset, 0); + struct thread *p,*th=arch_os_get_current_thread(); + 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 + * restarting */ + 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) { + 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); + } + } + } + + lock_ret = pthread_mutex_unlock(&all_threads_lock); + gc_assert(lock_ret == 0); +#ifdef LOCK_CREATE_THREAD + lock_ret = pthread_mutex_unlock(&create_thread_lock); + gc_assert(lock_ret == 0); +#endif + + FSHOW_SIGNAL((stderr,"/gc_start_the_world:end\n")); } +#endif -/* This is 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 unblock_sigcont_and_sleep(void) +int +thread_yield() { - sigset_t set; - sigemptyset(&set); - sigaddset(&set,SIGCONT); - sigwaitinfo(&set,0); - sigprocmask(SIG_UNBLOCK,&set,0); +#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 + } +}