X-Git-Url: http://repo.macrolet.net/gitweb/?a=blobdiff_plain;f=src%2Fruntime%2Fthread.c;h=a6b462fad6a6b61c31c978423b7441c3140b1d07;hb=d8e682fdfb7e8ba067e15aea0f3d1f8d37ca9eb1;hp=7d4520dd5e5154cdbdb70855f0cdfeeb33ed9822;hpb=aab24d77d0eeafb6880dce4420d3f422ef7a0971;p=sbcl.git diff --git a/src/runtime/thread.c b/src/runtime/thread.c index 7d4520d..a6b462f 100644 --- a/src/runtime/thread.c +++ b/src/runtime/thread.c @@ -1,77 +1,102 @@ #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 + #include "sbcl.h" +#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 */ struct thread *all_threads; -lispobj all_threads_lock; +volatile lispobj all_threads_lock; +volatile lispobj thread_start_lock; extern struct interrupt_data * global_interrupt_data; - -void get_spinlock(lispobj *word,int value); - -/* 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 - */ - -/* 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; +extern int linux_no_threads_p; 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; - /* wait here until our thread is linked into all_threads: see below */ - while(th->pid<1) sched_yield(); + 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 +void mark_thread_dead(struct thread *th) { + funcall1(SymbolFunction(HANDLE_THREAD_EXIT),alloc_number(th->os_thread)); + /* I hope it's safe for a thread to detach itself inside a + * cancellation cleanup */ + pthread_detach(th->os_thread); + th->state=STATE_DEAD; + /* FIXME: if gc hits here it will rip the stack from under us */ +} + +/* 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,ret; + function = th->unbound_marker; + th->unbound_marker = UNBOUND_MARKER_WIDETAG; + pthread_cleanup_push((void (*) (void *))mark_thread_dead,th); + 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; + ret = funcall0(function); + /* execute cleanup */ + pthread_cleanup_pop(1); + return ret; +} +#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 */ -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 */ 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 @@ -81,16 +106,15 @@ pid_t create_thread(lispobj initial_function) { BINDING_STACK_SIZE+ ALIEN_STACK_SIZE+ dynamic_values_bytes+ - 32*SIGSTKSZ - ); - if(!spaces) goto cleanup; + 32*SIGSTKSZ); + if(!spaces) + return NULL; per_thread=(union per_thread_data *) (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); @@ -114,12 +138,15 @@ pid_t create_thread(lispobj initial_function) { 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); +#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); @@ -128,17 +155,20 @@ pid_t create_thread(lispobj initial_function) { (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; + 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 +#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); #endif @@ -150,13 +180,13 @@ pid_t create_thread(lispobj initial_function) { * 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; @@ -169,7 +199,8 @@ pid_t create_thread(lispobj initial_function) { bind_variable(INTERRUPT_PENDING, NIL,th); bind_variable(INTERRUPTS_ENABLED,T,th); - th->interrupt_data=malloc(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, @@ -179,95 +210,252 @@ pid_t create_thread(lispobj initial_function) { 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) +{ + sigset_t newset,oldset; + sigemptyset(&newset); + sigaddset_blockable(&newset); + thread_sigmask(SIG_BLOCK, &newset, &oldset); + + get_spinlock(&all_threads_lock,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); - 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."); -#endif + protect_control_stack_guard_page(th->os_thread,1); + /* child will not start until this is set */ + th->os_thread=kid_tid; + release_spinlock(&all_threads_lock); - 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; + thread_sigmask(SIG_SETMASK,&oldset,0); } -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); +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 +os_thread_t create_thread(lispobj initial_function) { + struct thread *th; + os_thread_t kid_tid=0; + pthread_attr_t attr; + + if(linux_no_threads_p) return 0; + th=create_thread_struct(initial_function); + if(th==0) return 0; +#ifdef QSHOW_SIGNALS + SHOW("create_thread:waiting on lock"); +#endif + get_spinlock(&thread_start_lock,arch_os_get_current_thread()->os_thread); +#ifdef QSHOW_SIGNALS + SHOW("create_thread:got lock"); +#endif + /* The new thread inherits the restrictive signal mask set here, + * and enables signals again when it is set up properly. */ + { + sigset_t newset,oldset; + sigemptyset(&newset); + 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))) + kid_tid=0; + thread_sigmask(SIG_SETMASK,&oldset,0); + } + if(kid_tid>0) { + link_thread(th,kid_tid); + /* it's started and initialized, it's safe to gc */ + release_spinlock(&thread_start_lock); +#ifdef QSHOW_SIGNALS + SHOW("create_thread:released lock"); +#endif + /* by now the kid might have already exited */ + return kid_tid; + } else { + release_spinlock(&thread_start_lock); +#ifdef QSHOW_SIGNALS + SHOW("create_thread:released lock(failure)"); #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_invalidate((os_vm_address_t) th->control_stack_start, + ((sizeof (lispobj)) + * (th->control_stack_end-th->control_stack_start)) + + BINDING_STACK_SIZE+ALIEN_STACK_SIZE+dynamic_values_bytes+ + 32*SIGSTKSZ); + return 0; } - 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); } +#endif - -struct thread *find_thread_by_pid(pid_t pid) +struct thread *find_thread_by_os_thread(os_thread_t tid) { struct thread *th; for_each_thread(th) - if(th->pid==pid) return th; + if(th->os_thread==tid) return th; return 0; } +#if defined LISP_FEATURE_SB_THREAD +/* This is not needed unless #+SB-THREAD, as there's a trivial null + * unithread definition. */ +void reap_dead_threads() +{ + struct thread *th,*next,*prev=0; + th=all_threads; + while(th) { + next=th->next; + if(th->state==STATE_DEAD) { +#ifdef LISP_FEATURE_GENCGC + gc_alloc_update_page_tables(0, &th->alloc_region); +#endif + get_spinlock(&all_threads_lock,th->os_thread); + if(prev) prev->next=next; + else all_threads=next; + release_spinlock(&all_threads_lock); + if(th->tls_cookie>=0) arch_os_thread_cleanup(th); + os_invalidate((os_vm_address_t) th->control_stack_start, + ((sizeof (lispobj)) + * (th->control_stack_end-th->control_stack_start)) + + BINDING_STACK_SIZE+ALIEN_STACK_SIZE+dynamic_values_bytes+ + 32*SIGSTKSZ); + } else + prev=th; + th=next; + } +} -void block_sigcont(void) +int interrupt_thread(os_thread_t tid, lispobj function) { - /* 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 *th; + for_each_thread(th) + if((th->os_thread==tid) && (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. + */ + struct cons *c; + int kill_status; + /* mask the signals in case this thread is being interrupted */ + sigset_t newset,oldset; + sigemptyset(&newset); + sigaddset_blockable(&newset); + thread_sigmask(SIG_BLOCK, &newset, &oldset); + + get_spinlock(&th->interrupt_fun_lock, + (int)arch_os_get_current_thread()); + kill_status=thread_kill(th->os_thread,SIG_INTERRUPT_THREAD); + if(kill_status==0) { + c=alloc_cons(function,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; } -void unblock_sigcont_and_sleep(void) +/* 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. + */ + +void gc_stop_the_world() { - sigset_t set; - sigemptyset(&set); - sigaddset(&set,SIGCONT); - sigwaitinfo(&set,0); - sigprocmask(SIG_UNBLOCK,&set,0); + struct thread *p,*th=arch_os_get_current_thread(); +#ifdef QSHOW_SIGNALS + SHOW("gc_stop_the_world:begin"); +#endif + /* keep threads from starting while the world is stopped. */ + get_spinlock(&thread_start_lock,th->os_thread); +#ifdef QSHOW_SIGNALS + SHOW("gc_stop_the_world:locked"); +#endif + /* 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->os_thread!=0) && (p->state==STATE_RUNNING)) { + p->state=STATE_STOPPING; + if(thread_kill(p->os_thread,SIG_STOP_FOR_GC)==-1) { + /* FIXME: we can't kill the thread; assume because it died + * already */ + p->state=STATE_DEAD; + } + } + } +#ifdef QSHOW_SIGNALS + SHOW("gc_stop_the_world:signals sent"); +#endif + /* wait for the running threads to stop */ + for(p=all_threads;p;) { + if((p==th) || (p->os_thread==0) || (p->state==STATE_STARTING) || + (p->state==STATE_DEAD) || (p->state==STATE_STOPPED)) { + p=p->next; + } + } +#ifdef QSHOW_SIGNALS + SHOW("gc_stop_the_world:end"); +#endif } +void gc_start_the_world() +{ + struct thread *p,*th=arch_os_get_current_thread(); + /* 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; there can be threads just starting from before + * gc_stop_the_world, though */ +#ifdef QSHOW_SIGNALS + SHOW("gc_start_the_world:begin"); +#endif + for(p=all_threads;p;p=p->next) { + if((p!=th) && (p->os_thread!=0) && (p->state!=STATE_STARTING) && + (p->state!=STATE_DEAD)) { + if(p->state!=STATE_STOPPED) { + lose("gc_start_the_world: wrong thread state is %ld\n", + fixnum_value(p->state)); + } + thread_kill(p->os_thread,SIG_STOP_FOR_GC); + } + } + /* we must wait for all threads to leave stopped state else we + * risk signal accumulation and lose any meaning of + * thread->state */ + for(p=all_threads;p;) { + gc_assert(p->state!=STATE_STOPPING); + if((p==th) || (p->os_thread==0) || (p->state!=STATE_STOPPED)) { + p=p->next; + } + } + release_spinlock(&thread_start_lock); +#ifdef QSHOW_SIGNALS + SHOW("gc_start_the_world:end"); +#endif +} +#endif