0.9.2.9: thread objects
[sbcl.git] / src / runtime / thread.c
index a6b462f..6e3e978 100644 (file)
 int dynamic_values_bytes=4096*sizeof(lispobj); /* same for all threads */
 struct thread *all_threads;
 volatile lispobj all_threads_lock;
-volatile lispobj thread_start_lock;
 extern struct interrupt_data * global_interrupt_data;
 extern int linux_no_threads_p;
 
+/* When trying to get all_threads_lock one should make sure that
+ * sig_stop_for_gc is not blocked. Else there would be a possible
+ * deadlock: gc locks it, other thread blocks signals, gc sends stop
+ * request to other thread and waits, other thread blocks on lock. */
+void check_sig_stop_for_gc_can_arrive_or_lose()
+{
+    /* Get the current sigmask, by blocking the empty set. */
+    sigset_t empty,current;
+    sigemptyset(&empty);
+    thread_sigmask(SIG_BLOCK, &empty, &current);
+    if (sigismember(&current,SIG_STOP_FOR_GC))
+        lose("SIG_STOP_FOR_GC is blocked\n");
+    if (SymbolValue(INTERRUPTS_ENABLED,arch_os_get_current_thread()) == NIL)
+        lose("interrupts disabled\n");
+    if (arch_pseudo_atomic_atomic(NULL))
+        lose("n pseudo atomic\n");
+}
+
+#ifdef QSHOW_SIGNALS
+#define FSHOW_SIGNAL FSHOW
+#else
+#define FSHOW_SIGNAL(args)
+#endif
+
+#define GET_ALL_THREADS_LOCK(name) \
+    { \
+        sigset_t _newset,_oldset; \
+        sigemptyset(&_newset); \
+        sigaddset_blockable(&_newset); \
+        sigdelset(&_newset,SIG_STOP_FOR_GC); \
+        thread_sigmask(SIG_BLOCK, &_newset, &_oldset); \
+        check_sig_stop_for_gc_can_arrive_or_lose(); \
+        FSHOW_SIGNAL((stderr,"/%s:waiting on lock=%ld, thread=%ld\n",name, \
+               all_threads_lock,arch_os_get_current_thread()->os_thread)); \
+        get_spinlock(&all_threads_lock,(long)arch_os_get_current_thread()); \
+        FSHOW_SIGNAL((stderr,"/%s:got lock, thread=%ld\n", \
+               name,arch_os_get_current_thread()->os_thread));
+
+#define RELEASE_ALL_THREADS_LOCK(name) \
+        FSHOW_SIGNAL((stderr,"/%s:released lock\n",name)); \
+        release_spinlock(&all_threads_lock); \
+        thread_sigmask(SIG_SETMASK,&_oldset,0); \
+    }
+
 int
 initial_thread_trampoline(struct thread *th)
 {
@@ -54,14 +97,6 @@ initial_thread_trampoline(struct thread *th)
 }
 
 #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
@@ -71,26 +106,22 @@ void mark_thread_dead(struct thread *th) {
 int
 new_thread_trampoline(struct thread *th)
 {
-    lispobj function,ret;
+    lispobj function;
     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;   
+    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;
+    return funcall0(function);
 }
 #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 
+ * initialize all parts of it that can be initialized from another
+ * thread
  */
 
 struct thread * create_thread_struct(lispobj initial_function) {
@@ -123,7 +154,7 @@ struct thread * create_thread_struct(lispobj initial_function) {
        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) 
+       if(SymbolValue(FREE_TLS_INDEX,0)==UNBOUND_MARKER_WIDETAG)
            SetSymbolValue
                (FREE_TLS_INDEX,
                 make_fixnum(MAX_INTERRUPTS+
@@ -132,7 +163,7 @@ struct thread * create_thread_struct(lispobj initial_function) {
 #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);
@@ -176,7 +207,7 @@ struct thread * create_thread_struct(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 */
@@ -192,20 +223,20 @@ struct thread * create_thread_struct(lispobj initial_function) {
     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);
 
     th->interrupt_data = (struct interrupt_data *)
         os_validate(0,(sizeof (struct interrupt_data)));
-    if(all_threads) 
+    if(all_threads)
        memcpy(th->interrupt_data,
               arch_os_get_current_thread()->interrupt_data,
               sizeof (struct interrupt_data));
-    else 
+    else
        memcpy(th->interrupt_data,global_interrupt_data,
               sizeof (struct interrupt_data));
 
@@ -215,12 +246,6 @@ struct thread * create_thread_struct(lispobj initial_function) {
 
 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;
@@ -229,12 +254,9 @@ void link_thread(struct thread *th,os_thread_t kid_tid)
      * 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->os_thread,1);
+    protect_control_stack_guard_page(th,1);
     /* child will not start until this is set */
     th->os_thread=kid_tid;
-    release_spinlock(&all_threads_lock);
-
-    thread_sigmask(SIG_SETMASK,&oldset,0);
 }
 
 void create_initial_thread(lispobj initial_function) {
@@ -247,180 +269,185 @@ void create_initial_thread(lispobj initial_function) {
 }
 
 #ifdef LISP_FEATURE_SB_THREAD
-os_thread_t create_thread(lispobj initial_function) {
+
+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;
+    boolean r=1;
+    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)))
+        r=0;
+    thread_sigmask(SIG_SETMASK,&oldset,0);
+    return r;
+}
+
+struct thread *create_thread(lispobj initial_function) {
     struct thread *th;
     os_thread_t kid_tid=0;
-    pthread_attr_t attr;
+    boolean success;
 
     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) {
+
+    /* we must not be interrupted here after a successful
+     * create_os_thread, because the kid will be waiting for its
+     * thread struct to be linked */
+    GET_ALL_THREADS_LOCK("create_thread")
+
+    success=create_os_thread(th,&kid_tid);
+    if (success)
        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
+    else
        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;
-    }
-}
-#endif
 
-struct thread *find_thread_by_os_thread(os_thread_t tid) 
-{
-    struct thread *th;
-    for_each_thread(th)
-       if(th->os_thread==tid) return th;
-    return 0;
+    RELEASE_ALL_THREADS_LOCK("create_thread")
+
+    if (success)
+        return th;
+    else
+        return 0;
 }
+#endif
 
 #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() 
+/* called from lisp from the thread object finalizer */
+void reap_dead_thread(struct thread *th)
 {
-    struct thread *th,*next,*prev=0;
-    th=all_threads;
-    while(th) {
-       next=th->next;
-       if(th->state==STATE_DEAD) {
+    if(th->state!=STATE_DEAD)
+        lose("thread %lx is not joinable, state=%d\n",th,th->state);
 #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;
+    {
+        sigset_t newset,oldset;
+        sigemptyset(&newset);
+        sigaddset_blockable(&newset);
+        thread_sigmask(SIG_BLOCK, &newset, &oldset);
+        gc_alloc_update_page_tables(0, &th->alloc_region);
+        release_spinlock(&all_threads_lock);
+        thread_sigmask(SIG_SETMASK,&oldset,0);
     }
+#endif
+    GET_ALL_THREADS_LOCK("reap_dead_thread")
+    FSHOW((stderr,"/reap_dead_thread: reaping %ld\n",th->os_thread));
+    if(th->prev)
+        th->prev->next=th->next;
+    else all_threads=th->next;
+    if(th->next)
+        th->next->prev=th->prev;
+    RELEASE_ALL_THREADS_LOCK("reap_dead_thread")
+    if(th->tls_cookie>=0) arch_os_thread_cleanup(th);
+    gc_assert(pthread_join(th->os_thread,NULL)==0);
+    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);
 }
 
-int interrupt_thread(os_thread_t tid, lispobj function)
+int interrupt_thread(struct thread *th, lispobj function)
 {
-    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);
-        } 
+    /* A thread may also become dead after this test. */
+    if ((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=alloc_cons(function,NIL);
+        int kill_status;
+        /* interrupt_thread_handler locks this spinlock with
+         * interrupts blocked and it does so for the sake of
+         * arrange_return_to_lisp_function, so we must also block
+         * them. */
+        sigset_t newset,oldset;
+        sigemptyset(&newset);
+        sigaddset_blockable(&newset);
+        thread_sigmask(SIG_BLOCK, &newset, &oldset);
+        get_spinlock(&th->interrupt_fun_lock,
+                     (long)arch_os_get_current_thread());
+        kill_status=thread_kill(th->os_thread,SIG_INTERRUPT_THREAD);
+        if(kill_status==0) {
+            ((struct cons *)native_pointer(c))->cdr=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;
 }
 
-/* stopping the world is a two-stage process.  From this thread we signal 
+/* 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 
+ * 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.
  */
 
+/* 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 *p,*th=arch_os_get_current_thread();
-#ifdef QSHOW_SIGNALS
-    SHOW("gc_stop_the_world:begin");
-#endif
+    FSHOW_SIGNAL((stderr,"/gc_stop_the_world:waiting on lock, thread=%ld\n",
+                  th->os_thread));
     /* 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
+    get_spinlock(&all_threads_lock,(long)th);
+    FSHOW_SIGNAL((stderr,"/gc_stop_the_world:got lock, thread=%ld\n",
+                  th->os_thread));
     /* 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((p!=th) && (p->state==STATE_RUNNING)) {
+            FSHOW_SIGNAL((stderr,"/gc_stop_the_world:sending sig_stop to %ld\n",
+                          p->os_thread));
             if(thread_kill(p->os_thread,SIG_STOP_FOR_GC)==-1) {
-                /* FIXME: we can't kill the thread; assume because it died
-                 * already */
+                /* we can't kill the thread; assume because it died
+                 * since we last checked */
                 p->state=STATE_DEAD;
+                FSHOW_SIGNAL((stderr,"/gc_stop_the_world:assuming %ld dead\n",
+                   p->os_thread));
             }
         }
     }
-#ifdef QSHOW_SIGNALS
-    SHOW("gc_stop_the_world:signals sent");
-#endif
-    /* wait for the running threads to stop */
+    FSHOW_SIGNAL((stderr,"/gc_stop_the_world:signals sent\n"));
+    /* wait for the running threads to stop or finish */
     for(p=all_threads;p;) {
-        if((p==th) || (p->os_thread==0) || (p->state==STATE_STARTING) ||
-           (p->state==STATE_DEAD) || (p->state==STATE_STOPPED)) {
+        gc_assert(p->os_thread!=0);
+        gc_assert(p->state!=STATE_STARTING);
+        if((p==th) || (p->state==STATE_SUSPENDED) ||
+           (p->state==STATE_DEAD)) {
             p=p->next;
         }
     }
-#ifdef QSHOW_SIGNALS
-    SHOW("gc_stop_the_world:end");
-#endif
+    FSHOW_SIGNAL((stderr,"/gc_stop_the_world:end\n"));
 }
 
 void gc_start_the_world()
@@ -429,15 +456,12 @@ void gc_start_the_world()
     /* 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
+     * restarting */
+    FSHOW_SIGNAL((stderr,"/gc_start_the_world:begin\n"));
     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) {
+        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 %ld\n",
                      fixnum_value(p->state));
             }
@@ -448,14 +472,11 @@ void gc_start_the_world()
      * 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)) {
+        if((p==th) || (p->state!=STATE_SUSPENDED)) {
             p=p->next;
         }
     }
-    release_spinlock(&thread_start_lock);
-#ifdef QSHOW_SIGNALS
-    SHOW("gc_start_the_world:end");
-#endif
+    release_spinlock(&all_threads_lock);
+    FSHOW_SIGNAL((stderr,"/gc_start_the_world:end\n"));
 }
 #endif