Use safepoints for INTERRUPT-THREAD
[sbcl.git] / src / runtime / thread.c
index f594d56..d52c812 100644 (file)
@@ -33,7 +33,6 @@
 
 #include "runtime.h"
 #include "validate.h"           /* for BINDING_STACK_SIZE etc */
-#include "alloc.h"
 #include "thread.h"
 #include "arch.h"
 #include "target-arch-os.h"
 #include "genesis/cons.h"
 #include "genesis/fdefn.h"
 #include "interr.h"             /* for lose() */
+#include "alloc.h"
 #include "gc-internal.h"
+#include "cpputil.h"
+#include "pseudo-atomic.h"
+#include "interrupt.h"
+#include "lispregs.h"
 
 #ifdef LISP_FEATURE_WIN32
 /*
@@ -119,6 +123,80 @@ unlink_thread(struct thread *th)
     if (th->next)
         th->next->prev = th->prev;
 }
+
+/* Only access thread state with blockables blocked. */
+lispobj
+thread_state(struct thread *thread)
+{
+    lispobj state;
+    sigset_t old;
+    block_blockable_signals(NULL, &old);
+    os_sem_wait(thread->state_sem, "thread_state");
+    state = thread->state;
+    os_sem_post(thread->state_sem, "thread_state");
+    thread_sigmask(SIG_SETMASK, &old, NULL);
+    return state;
+}
+
+void
+set_thread_state(struct thread *thread, lispobj state)
+{
+    int i, waitcount = 0;
+    sigset_t old;
+    block_blockable_signals(NULL, &old);
+    os_sem_wait(thread->state_sem, "set_thread_state");
+    if (thread->state != state) {
+        if ((STATE_STOPPED==state) ||
+            (STATE_DEAD==state)) {
+            waitcount = thread->state_not_running_waitcount;
+            thread->state_not_running_waitcount = 0;
+            for (i=0; i<waitcount; i++)
+                os_sem_post(thread->state_not_running_sem, "set_thread_state (not running)");
+        }
+        if ((STATE_RUNNING==state) ||
+            (STATE_DEAD==state)) {
+            waitcount = thread->state_not_stopped_waitcount;
+            thread->state_not_stopped_waitcount = 0;
+            for (i=0; i<waitcount; i++)
+                os_sem_post(thread->state_not_stopped_sem, "set_thread_state (not stopped)");
+        }
+        thread->state = state;
+    }
+    os_sem_post(thread->state_sem, "set_thread_state");
+    thread_sigmask(SIG_SETMASK, &old, NULL);
+}
+
+void
+wait_for_thread_state_change(struct thread *thread, lispobj state)
+{
+    sigset_t old;
+    os_sem_t *wait_sem;
+    block_blockable_signals(NULL, &old);
+  start:
+    os_sem_wait(thread->state_sem, "wait_for_thread_state_change");
+    if (thread->state == state) {
+        switch (state) {
+        case STATE_RUNNING:
+            wait_sem = thread->state_not_running_sem;
+            thread->state_not_running_waitcount++;
+            break;
+        case STATE_STOPPED:
+            wait_sem = thread->state_not_stopped_sem;
+            thread->state_not_stopped_waitcount++;
+            break;
+        default:
+            lose("Invalid state in wait_for_thread_state_change: "OBJ_FMTX"\n", state);
+        }
+    } else {
+        wait_sem = NULL;
+    }
+    os_sem_post(thread->state_sem, "wait_for_thread_state_change");
+    if (wait_sem) {
+        os_sem_wait(wait_sem, "wait_for_thread_state_change");
+        goto start;
+    }
+    thread_sigmask(SIG_SETMASK, &old, NULL);
+}
 #endif
 
 static int
@@ -131,9 +209,16 @@ initial_thread_trampoline(struct thread *th)
 #ifdef LISP_FEATURE_SB_THREAD
     pthread_setspecific(lisp_thread, (void *)1);
 #endif
+#if defined(THREADS_USING_GCSIGNAL) && 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;
+#ifdef LISP_FEATURE_SB_SAFEPOINT
+    pthread_mutex_lock(thread_qrl(th));
+#endif
     link_thread(th);
     th->os_thread=thread_self();
 #ifndef LISP_FEATURE_WIN32
@@ -153,20 +238,6 @@ initial_thread_trampoline(struct thread *th)
 }
 
 #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
@@ -266,7 +337,9 @@ new_thread_trampoline(struct thread *th)
 
     FSHOW((stderr,"/creating thread %lu\n", thread_self()));
     check_deferrables_blocked_or_lose(0);
+#ifndef LISP_FEATURE_SB_SAFEPOINT
     check_gc_signals_unblocked_or_lose(0);
+#endif
     pthread_setspecific(lisp_thread, (void *)1);
     function = th->no_tls_value_marker;
     th->no_tls_value_marker = NO_TLS_VALUE_MARKER_WIDETAG;
@@ -283,12 +356,33 @@ new_thread_trampoline(struct thread *th)
      * 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). */
+#ifdef LISP_FEATURE_SB_SAFEPOINT
+    *th->csp_around_foreign_call = (lispobj)&function;
+    pthread_mutex_lock(thread_qrl(th));
+#endif
     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);
 
+    /* Kludge: Changed the order of some steps between the safepoint/
+     * non-safepoint versions of this code.  Can we unify this more?
+     */
+#ifdef LISP_FEATURE_SB_SAFEPOINT
+    WITH_GC_AT_SAFEPOINTS_ONLY() {
+        result = funcall0(function);
+        block_blockable_signals(0, 0);
+        gc_alloc_update_page_tables(BOXED_PAGE_FLAG, &th->alloc_region);
+    }
+    lock_ret = pthread_mutex_lock(&all_threads_lock);
+    gc_assert(lock_ret == 0);
+    unlink_thread(th);
+    lock_ret = pthread_mutex_unlock(&all_threads_lock);
+    gc_assert(lock_ret == 0);
+    pthread_mutex_unlock(thread_qrl(th));
+    set_thread_state(th,STATE_DEAD);
+#else
     result = funcall0(function);
 
     /* Block GC */
@@ -304,23 +398,18 @@ new_thread_trampoline(struct thread *th)
     unlink_thread(th);
     pthread_mutex_unlock(&all_threads_lock);
     gc_assert(lock_ret == 0);
+#endif
 
     if(th->tls_cookie>=0) arch_os_thread_cleanup(th);
-    pthread_mutex_destroy(th->state_lock);
-    pthread_cond_destroy(th->state_cond);
+    os_sem_destroy(th->state_sem);
+    os_sem_destroy(th->state_not_running_sem);
+    os_sem_destroy(th->state_not_stopped_sem);
 
     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));
+    mach_lisp_thread_destroy(th);
 #endif
 
     schedule_thread_post_mortem(th);
@@ -340,6 +429,12 @@ free_thread_struct(struct thread *th)
                   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
@@ -371,35 +466,37 @@ create_thread_struct(lispobj initial_function) {
     aligned_spaces = (void *)((((unsigned long)(char *)spaces)
                                + THREAD_ALIGNMENT_BYTES-1)
                               &~(unsigned long)(THREAD_ALIGNMENT_BYTES-1));
-    per_thread=(union per_thread_data *)
+    void* csp_page=
         (aligned_spaces+
          thread_control_stack_size+
          BINDING_STACK_SIZE+
-         ALIEN_STACK_SIZE +
-         THREAD_STATE_LOCK_SIZE);
+         ALIEN_STACK_SIZE);
+    per_thread=(union per_thread_data *)
+        (csp_page + THREAD_CSP_PAGE_SIZE);
+    struct nonpointer_thread_data *nonpointer_data
+        = (void *) &per_thread->dynamic_values[TLS_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
@@ -413,19 +510,36 @@ create_thread_struct(lispobj initial_function) {
     th->binding_stack_start=
         (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_SAFEPOINT
+    th->pc_around_foreign_call = 0;
+    th->csp_around_foreign_call = csp_page;
+#endif
+
 #ifdef LISP_FEATURE_SB_THREAD
+    /* Contrary to the "allocate all the spaces at once" comment above,
+     * the os_attr is allocated separately.  We cannot put it into the
+     * nonpointer data, because it's used for post_mortem and freed
+     * separately */
     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);
+    th->nonpointer_data = nonpointer_data;
+    th->state_sem=&nonpointer_data->state_sem;
+    th->state_not_running_sem=&nonpointer_data->state_not_running_sem;
+    th->state_not_stopped_sem=&nonpointer_data->state_not_stopped_sem;
+    th->state_not_running_waitcount = 0;
+    th->state_not_stopped_waitcount = 0;
+    os_sem_init(th->state_sem, 1);
+    os_sem_init(th->state_not_running_sem, 0);
+    os_sem_init(th->state_not_stopped_sem, 0);
+# ifdef LISP_FEATURE_SB_SAFEPOINT
+    pthread_mutex_init(thread_qrl(th), NULL);
+# endif
 #endif
     th->state=STATE_RUNNING;
 #ifdef LISP_FEATURE_STACK_GROWS_DOWNWARD_NOT_UPWARD
@@ -434,12 +548,22 @@ 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
@@ -452,12 +576,8 @@ 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);
@@ -468,9 +588,23 @@ create_thread_struct(lispobj initial_function) {
     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
+#if defined(LISP_FEATURE_SB_SAFEPOINT)
+    bind_variable(GC_SAFE,NIL,th);
+    bind_variable(IN_SAFEPOINT,NIL,th);
+#endif
+#ifdef LISP_FEATURE_SB_THRUPTION
+    bind_variable(THRUPTION_PENDING,NIL,th);
+    bind_variable(RESTART_CLUSTERS,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)));
@@ -489,21 +623,12 @@ create_thread_struct(lispobj initial_function) {
     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);
-
-#endif
-
 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
-        setup_mach_exception_handling_thread();
-#endif
         initial_thread_trampoline(th); /* no return */
     } else lose("can't create initial thread\n");
 }
@@ -586,6 +711,10 @@ os_thread_t create_thread(lispobj initial_function) {
  * 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.
  */
+/*
+ * (With SB-SAFEPOINT, see the definitions in safepoint.c instead.)
+ */
+#ifndef LISP_FEATURE_SB_SAFEPOINT
 
 /* 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
@@ -657,7 +786,7 @@ void gc_start_the_world()
         if (p!=th) {
             lispobj state = thread_state(p);
             if (state != STATE_DEAD) {
-                if(state != STATE_SUSPENDED) {
+                if(state != STATE_STOPPED) {
                     lose("gc_start_the_world: wrong thread state is %d\n",
                          fixnum_value(state));
                 }
@@ -677,7 +806,9 @@ void gc_start_the_world()
 
     FSHOW_SIGNAL((stderr,"/gc_start_the_world:end\n"));
 }
-#endif
+
+#endif /* !LISP_FEATURE_SB_SAFEPOINT */
+#endif /* !LISP_FEATURE_SB_THREAD */
 
 int
 thread_yield()
@@ -689,6 +820,19 @@ thread_yield()
 #endif
 }
 
+int
+wake_thread(os_thread_t os_thread)
+{
+#ifdef LISP_FEATURE_WIN32
+# define SIGPIPE 1
+#endif
+#if !defined(LISP_FEATURE_SB_THRUPTION) || defined(LISP_FEATURE_WIN32)
+    return kill_safely(os_thread, SIGPIPE);
+#else
+    return wake_thread_posix(os_thread);
+#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
@@ -698,13 +842,13 @@ thread_yield()
  * (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. */
+ * exiting.
+ *
+ * Note (DFL, 2011-06-22): At the time of writing, this function is only
+ * used for INTERRUPT-THREAD, hence the wake_thread special-case for
+ * Windows is OK. */
 int
 kill_safely(os_thread_t os_thread, int signal)
 {
@@ -713,6 +857,23 @@ kill_safely(os_thread_t os_thread, int signal)
 #ifdef LISP_FEATURE_SB_THREAD
         sigset_t oldset;
         struct thread *thread;
+        /* Frequent special case: resignalling to self.  The idea is
+         * that leave_region safepoint will acknowledge the signal, so
+         * there is no need to take locks, roll thread to safepoint
+         * etc. */
+        /* Kludge (on safepoint builds): At the moment, this isn't just
+         * an optimization; rather it masks the fact that
+         * gc_stop_the_world() grabs the all_threads mutex without
+         * releasing it, and since we're not using recursive pthread
+         * mutexes, the pthread_mutex_lock() around the all_threads loop
+         * would go wrong.  Why are we running interruptions while
+         * stopping the world though?  Test case is (:ASYNC-UNWIND
+         * :SPECIALS), especially with s/10/100/ in both loops. */
+        if (os_thread == pthread_self()) {
+            pthread_kill(os_thread, signal);
+            return 0;
+        }
+
         /* pthread_kill is not async signal safe and we don't want to be
          * interrupted while holding the lock. */
         block_deferrable_signals(0, &oldset);