X-Git-Url: http://repo.macrolet.net/gitweb/?a=blobdiff_plain;f=src%2Fruntime%2Fthread.c;h=3092fc078e80919e0c00d98ce86c41b86e7c6ae3;hb=b85552a9ed94019afd70b11b069f3c9b0afdaaa8;hp=ddcc8fdb3bb8f59b11ad5c67fc2d2a8280b7bcc0;hpb=e365f2f7a9c66d307b48fee70778f4eaa84bdcc0;p=sbcl.git diff --git a/src/runtime/thread.c b/src/runtime/thread.c index ddcc8fd..3092fc0 100644 --- a/src/runtime/thread.c +++ b/src/runtime/thread.c @@ -1,7 +1,9 @@ #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 @@ -13,16 +15,14 @@ #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 */ 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 int countdown_to_gc; extern struct interrupt_data * global_interrupt_data; void get_spinlock(lispobj *word,int value); @@ -50,15 +50,20 @@ new_thread_trampoline(struct thread *th) fprintf(stderr, "/continue\n"); } th->unbound_marker = UNBOUND_MARKER_WIDETAG; +#ifdef LISP_FEATURE_SB_THREAD /* wait here until our thread is linked into all_threads: see below */ while(th->pid<1) sched_yield(); +#else + if(th->pid < 1) + lose("th->pid not set up right"); +#endif if(arch_os_thread_init(th)==0) return 1; /* failure. no, really */ -#ifdef LISP_FEATURE_SB_THREAD - return call_into_lisp(function,args,0); -#else +#if !defined(LISP_FEATURE_SB_THREAD) && defined(LISP_FEATURE_X86) return call_into_lisp_first_time(function,args,0); +#else + return funcall0(function); #endif } @@ -95,6 +100,7 @@ pid_t create_thread(lispobj initial_function) { 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; @@ -104,7 +110,6 @@ pid_t create_thread(lispobj initial_function) { make_fixnum(MAX_INTERRUPTS+ sizeof(struct thread)/sizeof(lispobj)), 0); -#ifdef LISP_FEATURE_SB_THREAD #define STATIC_TLS_INIT(sym,field) \ ((struct symbol *)(sym-OTHER_POINTER_LOWTAG))->tls_index= \ make_fixnum(THREAD_SLOT_OFFSET_WORDS(field)) @@ -112,9 +117,12 @@ pid_t create_thread(lispobj initial_function) { 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); +#endif #undef STATIC_TLS_INIT #endif } @@ -122,22 +130,28 @@ pid_t create_thread(lispobj initial_function) { th->control_stack_start = spaces; th->binding_stack_start= (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); th->binding_stack_pointer=th->binding_stack_start; th->this=th; th->pid=0; + 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 */ #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); +#endif +#ifdef LISP_FEATURE_GENCGC gc_set_region_empty(&th->alloc_region); +#endif #ifndef LISP_FEATURE_SB_THREAD /* the tls-points-into-struct-thread trick is only good for threaded @@ -147,22 +161,28 @@ pid_t create_thread(lispobj initial_function) { * 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(BINDING_STACK_POINTER,th->binding_stack_pointer,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); +#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); bind_variable(CURRENT_UNWIND_PROTECT_BLOCK,make_fixnum(0),th); bind_variable(FREE_INTERRUPT_CONTEXT_INDEX,make_fixnum(0),th); bind_variable(INTERRUPT_PENDING, NIL,th); bind_variable(INTERRUPTS_ENABLED,T,th); - th->interrupt_data=malloc(sizeof (struct interrupt_data)); + th->interrupt_data=os_validate(0,(sizeof (struct interrupt_data))); if(all_threads) - memcpy(th->interrupt_data,arch_os_get_current_thread()->interrupt_data, + memcpy(th->interrupt_data, + arch_os_get_current_thread()->interrupt_data, sizeof (struct interrupt_data)); else memcpy(th->interrupt_data,global_interrupt_data, @@ -191,7 +211,7 @@ pid_t create_thread(lispobj initial_function) { * 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; + release_spinlock(&all_threads_lock); th->pid=kid_pid; /* child will not start until this is set */ #ifndef LISP_FEATURE_SB_THREAD new_thread_trampoline(all_threads); /* call_into_lisp */ @@ -211,8 +231,12 @@ void destroy_thread (struct thread *th) { /* precondition: the unix task has already been killed and exited. * This is called by the parent */ +#ifdef LISP_FEATURE_GENCGC gc_alloc_update_page_tables(0, &th->alloc_region); +#endif get_spinlock(&all_threads_lock,th->pid); + if(countdown_to_gc>0) countdown_to_gc--; + th->state=STATE_STOPPED; if(th==all_threads) all_threads=th->next; else { @@ -220,11 +244,12 @@ void destroy_thread (struct thread *th) while(th1->next!=th) th1=th1->next; th1->next=th->next; /* unlink */ } - all_threads_lock=0; + release_spinlock(&all_threads_lock); if(th && th->tls_cookie>=0) arch_os_thread_cleanup(th); os_invalidate((os_vm_address_t) th->control_stack_start, - THREAD_CONTROL_STACK_SIZE+BINDING_STACK_SIZE+ - ALIEN_STACK_SIZE+dynamic_values_bytes+ + ((sizeof (lispobj)) + * (th->control_stack_end-th->control_stack_start)) + + BINDING_STACK_SIZE+ALIEN_STACK_SIZE+dynamic_values_bytes+ 32*SIGSTKSZ); } @@ -237,18 +262,10 @@ struct thread *find_thread_by_pid(pid_t pid) return 0; } - -void get_spinlock(lispobj *word,int value) -{ - u32 eax=0; - do { - asm ("xor %0,%0;cmpxchg %1,%2" - : "=a" (eax) - : "r" (value), "m" (*word) - : "memory", "cc"); - } while(eax!=0); -} - +/* These are not needed unless #+SB-THREAD, and since sigwaitinfo() + * doesn't seem to be easily available everywhere (OpenBSD...) it's + * more trouble than it's worth to compile it when not needed. */ +#if defined LISP_FEATURE_SB_THREAD void block_sigcont(void) { /* don't allow ourselves to receive SIGCONT while we're in the @@ -256,7 +273,7 @@ void block_sigcont(void) */ sigset_t newset; sigemptyset(&newset); - sigaddset(&newset,SIGCONT); + sigaddset(&newset,SIG_DEQUEUE); sigprocmask(SIG_BLOCK, &newset, 0); } @@ -264,8 +281,79 @@ void unblock_sigcont_and_sleep(void) { sigset_t set; sigemptyset(&set); - sigaddset(&set,SIGCONT); - sigwaitinfo(&set,0); + sigaddset(&set,SIG_DEQUEUE); + do { + errno=0; + sigwaitinfo(&set,0); + }while(errno==EINTR); sigprocmask(SIG_UNBLOCK,&set,0); } +int interrupt_thread(pid_t pid, lispobj function) +{ + union sigval sigval; + sigval.sival_int=function; + + return sigqueue(pid, SIG_INTERRUPT_THREAD, sigval); +} + +/* stopping the world is a two-stage process. From this thread we signal + * all the others with SIG_STOP_FOR_GC. The handler for this thread does + * the usual pseudo-atomic checks (we don't want to stop a thread while + * it's in the middle of allocation) then kills _itself_ with SIGSTOP. + * At any given time, countdown_to_gc should reflect the number of threads + * signalled but which haven't yet come to rest + */ + +void gc_stop_the_world() +{ + /* stop all other threads by sending them SIG_STOP_FOR_GC */ + struct thread *p,*th=arch_os_get_current_thread(); + struct thread *tail=0; + int finished=0; + do { + get_spinlock(&all_threads_lock,th->pid); + if(tail!=all_threads) { + /* new threads always get consed onto the front of all_threads, + * and may be created by any thread that we haven't signalled + * yet or hasn't received our signal and stopped yet. So, check + * for them on each time around */ + for(p=all_threads;p!=tail;p=p->next) { + if(p==th) continue; + /* if the head of all_threads is removed during + * gc_stop_the_world, we may take a second trip through the + * list and end up counting twice as many threads to wait for + * as actually exist */ + if(p->state!=STATE_RUNNING) continue; + countdown_to_gc++; + p->state=STATE_STOPPING; + /* Note no return value check from kill(). If the + * thread had been reaped already, we kill it and + * increment countdown_to_gc anyway. This is to avoid + * complicating the logic in destroy_thread, which would + * otherwise have to know whether the thread died before or + * after it was killed + */ + kill(p->pid,SIG_STOP_FOR_GC); + } + tail=all_threads; + } else { + finished=(countdown_to_gc==0); + } + release_spinlock(&all_threads_lock); + sched_yield(); + } while(!finished); +} + +void gc_start_the_world() +{ + struct thread *p,*th=arch_os_get_current_thread(); + get_spinlock(&all_threads_lock,th->pid); + for(p=all_threads;p;p=p->next) { + if(p==th) continue; + p->state=STATE_RUNNING; + kill(p->pid,SIGCONT); + } + release_spinlock(&all_threads_lock); +} +#endif