#include <sys/wait.h>
#endif
+#ifdef LISP_FEATURE_MACH_EXCEPTION_HANDLER
+#include <mach/mach.h>
+#include <mach/mach_error.h>
+#include <mach/mach_types.h>
+#endif
+
#include "runtime.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"
#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
+
#define ALIEN_STACK_SIZE (1*1024*1024) /* 1Mb size chosen at random */
-struct freeable_stack {
+#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;
- os_vm_address_t stack;
+ pthread_attr_t *os_attr;
+ os_vm_address_t os_address;
};
-static struct freeable_stack * volatile freeable_stack = 0;
+#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=4096*sizeof(lispobj); /* same for all threads */
+int dynamic_values_bytes=TLS_SIZE*sizeof(lispobj); /* same for all threads */
struct thread * volatile all_threads;
extern struct interrupt_data * global_interrupt_data;
-extern int linux_no_threads_p;
#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
#if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
#endif
}
-#define THREAD_STRUCT_SIZE (THREAD_CONTROL_STACK_SIZE + BINDING_STACK_SIZE + \
- ALIEN_STACK_SIZE + dynamic_values_bytes + \
- 32 * SIGSTKSZ)
+#define THREAD_STRUCT_SIZE (thread_control_stack_size + BINDING_STACK_SIZE + \
+ ALIEN_STACK_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
-free_thread_stack_later(struct thread *thread_to_be_cleaned_up)
+perform_thread_post_mortem(struct thread_post_mortem *post_mortem)
{
- struct freeable_stack *new_freeable_stack = 0;
- if (thread_to_be_cleaned_up) {
- new_freeable_stack = (struct freeable_stack *)
- os_validate(0, sizeof(struct freeable_stack));
- new_freeable_stack->os_thread = thread_to_be_cleaned_up->os_thread;
- new_freeable_stack->stack = (os_vm_address_t)
- thread_to_be_cleaned_up->control_stack_start;
+#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);
}
- new_freeable_stack = (struct freeable_stack *)
- swap_lispobjs((lispobj *)(void *)&freeable_stack,
- (lispobj)new_freeable_stack);
- if (new_freeable_stack) {
- FSHOW((stderr,"/reaping %lu\n", new_freeable_stack->os_thread));
- /* Under NPTL pthread_join really waits until the thread
- * exists and the stack can be safely freed. This is sadly not
- * mandated by the pthread spec. */
- gc_assert(pthread_join(new_freeable_stack->os_thread, NULL) == 0);
- os_invalidate(new_freeable_stack->stack, THREAD_STRUCT_SIZE);
- os_invalidate((os_vm_address_t) new_freeable_stack,
- sizeof(struct freeable_stack));
+}
+
+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
}
}
new_thread_trampoline(struct thread *th)
{
lispobj function;
- int result;
+ int result, lock_ret;
+
FSHOW((stderr,"/creating thread %lu\n", thread_self()));
function = th->no_tls_value_marker;
th->no_tls_value_marker = NO_TLS_VALUE_MARKER_WIDETAG;
* 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). */
- pthread_mutex_lock(&all_threads_lock);
+ lock_ret = pthread_mutex_lock(&all_threads_lock);
+ gc_assert(lock_ret == 0);
link_thread(th);
- pthread_mutex_unlock(&all_threads_lock);
+ lock_ret = pthread_mutex_unlock(&all_threads_lock);
+ gc_assert(lock_ret == 0);
result = funcall0(function);
+
+ /* Block GC */
+ block_blockable_signals();
th->state=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. */
- pthread_mutex_lock(&all_threads_lock);
- gc_alloc_update_page_tables(0, &th->alloc_region);
+ 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);
os_invalidate((os_vm_address_t)th->interrupt_data,
(sizeof (struct interrupt_data)));
- free_thread_stack_later(th);
- FSHOW((stderr,"/exiting thread %lu\n", thread_self()));
+
+#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 %p\n", thread_self()));
return result;
}
if (th->interrupt_data)
os_invalidate((os_vm_address_t) th->interrupt_data,
(sizeof (struct interrupt_data)));
- os_invalidate((os_vm_address_t) th->control_stack_start,
+ os_invalidate((os_vm_address_t) th->os_address,
THREAD_STRUCT_SIZE);
}
union per_thread_data *per_thread;
struct thread *th=0; /* subdue gcc */
void *spaces=0;
+ void *aligned_spaces=0;
#ifdef LISP_FEATURE_SB_THREAD
- int i;
+ 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 */
+ * 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;
+ 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+
+ (aligned_spaces+
+ thread_control_stack_size+
BINDING_STACK_SIZE+
ALIEN_STACK_SIZE);
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_ATOMIC,pseudo_atomic_atomic);
- STATIC_TLS_INIT(PSEUDO_ATOMIC_INTERRUPTED,pseudo_atomic_interrupted);
+ STATIC_TLS_INIT(PSEUDO_ATOMIC_BITS,pseudo_atomic_bits);
#endif
#undef STATIC_TLS_INIT
}
#endif
th=&per_thread->thread;
- th->control_stack_start = spaces;
+ 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);
th->binding_stack_pointer=th->binding_stack_start;
th->this=th;
th->os_thread=0;
+#ifdef LISP_FEATURE_SB_THREAD
+ th->os_attr=malloc(sizeof(pthread_attr_t));
+#endif
th->state=STATE_RUNNING;
#ifdef LISP_FEATURE_STACK_GROWS_DOWNWARD_NOT_UPWARD
th->alien_stack_pointer=((void *)th->alien_stack_start
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;
- th->pseudo_atomic_atomic=0;
+ th->pseudo_atomic_bits=0;
#endif
#ifdef LISP_FEATURE_GENCGC
gc_set_region_empty(&th->alloc_region);
#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);
+ 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;
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->pending_handler = 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);
+
+#endif
+
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
initial_thread_trampoline(th); /* no return */
} else lose("can't create initial thread\n");
}
{
/* 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;
+ int retcode = 0, initcode;
+
+ FSHOW_SIGNAL((stderr,"/create_os_thread: creating new thread\n"));
+
+#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
sigemptyset(&newset);
/* Blocking deferrable signals is enough, no need to block
* SIG_STOP_FOR_GC because the child process is not linked onto
sigaddset_deferrable(&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)))
+ 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, printf("pthread_create returned %d, errno %d\n", retcode, errno)));
+ FSHOW_SIGNAL((stderr, "wanted stack size %d, min stack size %d\n",
+ cstack_size, PTHREAD_STACK_MIN));
+ if(retcode < 0) {
+ perror("create_os_thread");
+ }
r=0;
+ }
+
thread_sigmask(SIG_SETMASK,&oldset,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
return r;
}
struct thread *th;
os_thread_t kid_tid;
- if(linux_no_threads_p) return 0;
-
/* 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
/* Send the signo to os_thread, retry if the rt signal queue is
* full. */
-static int kill_thread_safely(os_thread_t os_thread, int signo)
+int
+kill_thread_safely(os_thread_t os_thread, int signo)
{
int r;
/* The man page does not mention EAGAIN as a valid return value
void gc_stop_the_world()
{
struct thread *p,*th=arch_os_get_current_thread();
- int status;
+ 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, thread=%lu\n",
+ th->os_thread));
+ lock_ret = pthread_mutex_lock(&create_thread_lock);
+ gc_assert(lock_ret == 0);
+ FSHOW_SIGNAL((stderr,"/gc_stop_the_world:got create_thread_lock, thread=%lu\n",
+ th->os_thread));
+#endif
FSHOW_SIGNAL((stderr,"/gc_stop_the_world:waiting on lock, thread=%lu\n",
th->os_thread));
/* keep threads from starting while the world is stopped. */
- pthread_mutex_lock(&all_threads_lock); \
+ lock_ret = pthread_mutex_lock(&all_threads_lock); \
+ gc_assert(lock_ret == 0);
+
FSHOW_SIGNAL((stderr,"/gc_stop_the_world:got lock, thread=%lu\n",
th->os_thread));
/* 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: p->state: %x\n", p->state));
if((p!=th) && ((p->state==STATE_RUNNING))) {
- FSHOW_SIGNAL((stderr,"/gc_stop_the_world: suspending %lu\n",
- p->os_thread));
+ FSHOW_SIGNAL((stderr,"/gc_stop_the_world: suspending %x, os_thread %x\n",
+ p, p->os_thread));
status=kill_thread_safely(p->os_thread,SIG_STOP_FOR_GC);
if (status==ESRCH) {
/* This thread has exited. */
FSHOW_SIGNAL((stderr,"/gc_stop_the_world:signals sent\n"));
/* wait for the running threads to stop or finish */
for(p=all_threads;p;) {
+ FSHOW_SIGNAL((stderr,"/gc_stop_the_world: th: %p, p: %p\n", th, p));
if((p!=th) && (p->state==STATE_RUNNING)) {
sched_yield();
} else {
void gc_start_the_world()
{
struct thread *p,*th=arch_os_get_current_thread();
- int status;
+ int status, 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
FSHOW_SIGNAL((stderr, "/gc_start_the_world: resuming %lu\n",
p->os_thread));
p->state=STATE_RUNNING;
+
+#if defined(SIG_RESUME_FROM_GC)
+ status=kill_thread_safely(p->os_thread,SIG_RESUME_FROM_GC);
+#else
status=kill_thread_safely(p->os_thread,SIG_STOP_FOR_GC);
+#endif
if (status) {
lose("cannot resume thread=%lu: %d, %s\n",
p->os_thread,status,strerror(status));
* SIG_STOP_FOR_GC wouldn't need to be a rt signal. That has some
* performance implications, but does away with the 'rt signal
* queue full' problem. */
- pthread_mutex_unlock(&all_threads_lock); \
+
+ 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
+
+int
+thread_yield()
+{
+#ifdef LISP_FEATURE_SB_THREAD
+ return sched_yield();
+#else
+ return 0;
+#endif
+}