Libc-825.25.tar.gz

[apple/libc.git] / pthreads / pthread.c

diff --git a/pthreads/pthread.c b/pthreads/pthread.c
index 5e9aefbfc05458b3db2775cc267bb44f4578a744..d6d116c3ed23eee22d642b39e424d16b6ce3da1e 100644 (file)
--- a/pthreads/pthread.c
+++ b/pthreads/pthread.c
@@ -61,14 +61,17 @@
 #include <sys/resource.h>
 #include <sys/sysctl.h>
 #include <sys/queue.h>
 #include <sys/resource.h>
 #include <sys/sysctl.h>
 #include <sys/queue.h>

+#include <sys/mman.h>

 #include <machine/vmparam.h>
 #include <mach/vm_statistics.h>
 #include <machine/vmparam.h>
 #include <mach/vm_statistics.h>

+#include <mach/mach_init.h>

 #define        __APPLE_API_PRIVATE
 #include <machine/cpu_capabilities.h>
 #include <libkern/OSAtomic.h>
 #if defined(__ppc__)
 #include <libkern/OSCrossEndian.h>
 #endif
 #define        __APPLE_API_PRIVATE
 #include <machine/cpu_capabilities.h>
 #include <libkern/OSAtomic.h>
 #if defined(__ppc__)
 #include <libkern/OSCrossEndian.h>
 #endif

+#include <dispatch/private.h> /* for at_fork handlers */

 
 
 extern int _pthread_setcancelstate_internal(int state, int *oldstate, int conforming);
 
 
 extern int _pthread_setcancelstate_internal(int state, int *oldstate, int conforming);


@@ -88,13 +91,15 @@ extern void mig_init(int);
 static int _pthread_create_pthread_onstack(pthread_attr_t *attrs, void **stack, pthread_t *thread);
 static kern_return_t _pthread_free_pthread_onstack(pthread_t t, int freestruct, int termthread);
 static void _pthread_struct_init(pthread_t t, const pthread_attr_t *attrs, void * stack, size_t stacksize, int kernalloc, int nozero);
 static int _pthread_create_pthread_onstack(pthread_attr_t *attrs, void **stack, pthread_t *thread);
 static kern_return_t _pthread_free_pthread_onstack(pthread_t t, int freestruct, int termthread);
 static void _pthread_struct_init(pthread_t t, const pthread_attr_t *attrs, void * stack, size_t stacksize, int kernalloc, int nozero);

-static void _pthread_tsd_reinit(pthread_t t);

 static int  _new_pthread_create_suspended(pthread_t *thread, 
               const pthread_attr_t *attr,
               void *(*start_routine)(void *), 
               void *arg,
                int create_susp);
 
 static int  _new_pthread_create_suspended(pthread_t *thread, 
               const pthread_attr_t *attr,
               void *(*start_routine)(void *), 
               void *arg,
                int create_susp);
 

+/* the registered libdispatch worker function */
+static void (*__libdispatch_workerfunction)(int, int, void *) = NULL;
+

 /* Get CPU capabilities from the kernel */
 __private_extern__ void _init_cpu_capabilities(void);
 
 /* Get CPU capabilities from the kernel */
 __private_extern__ void _init_cpu_capabilities(void);
 


@@ -113,6 +118,47 @@ typedef struct _pthread_reap_msg_t {
        mach_msg_trailer_t trailer;
 } pthread_reap_msg_t;
 
        mach_msg_trailer_t trailer;
 } pthread_reap_msg_t;
 

+/* Utilitie */
+
+__private_extern__ uintptr_t commpage_pfz_base=0;
+
+void __pthread_pfz_setup(const char *apple[]) __attribute__ ((visibility ("hidden")));
+
+static uintptr_t __pfz_from_kernel(const char *str)
+{
+       unsigned long tmpval;
+       /* Skip over key to the first value */
+       str = strchr(str, '=');
+       if (str == NULL)
+               return 0;
+       str++;
+       tmpval = strtoul(str, NULL, 0); /* may err by 0 or ULONG_MAX */
+       if (tmpval == ULONG_MAX)
+               tmpval = 0;
+
+       return (uintptr_t) tmpval;
+}
+
+void
+__pthread_pfz_setup(const char *apple[])
+{
+       const char **p;
+       for (p = apple; p && *p; p++) {
+               /* checking if matching apple variable is at begining */
+               if (strstr(*p, "pfz=") == *p) {
+                       commpage_pfz_base = __pfz_from_kernel(*p);
+                       bzero(*p,strlen(*p));
+                       break;
+               }
+       }
+
+       if (commpage_pfz_base == 0)
+               commpage_pfz_base = _COMM_PAGE_TEXT_START;
+
+       return;
+}
+
+

 /* We'll implement this when the main thread is a pthread */
 /* Use the local _pthread struct to avoid malloc before our MiG reply port is set */
 static struct _pthread _thread = {0};
 /* We'll implement this when the main thread is a pthread */
 /* Use the local _pthread struct to avoid malloc before our MiG reply port is set */
 static struct _pthread _thread = {0};


@@ -124,6 +170,8 @@ int __is_threaded = 0;
 /* _pthread_count is protected by _pthread_list_lock */
 static int _pthread_count = 1;
 int __unix_conforming = 0;
 /* _pthread_count is protected by _pthread_list_lock */
 static int _pthread_count = 1;
 int __unix_conforming = 0;

+static int __workqueue_newspis = 0;
+static int __workqueue_oldspis = 0;

 __private_extern__ size_t pthreadsize = 0;
 
 /* under rosetta we will use old style creation of threads */
 __private_extern__ size_t pthreadsize = 0;
 
 /* under rosetta we will use old style creation of threads */


@@ -145,7 +193,7 @@ __private_extern__ void _spin_lock_retry(pthread_lock_t *lock)
        } while(!_spin_lock_try(lock));
 }
 
        } while(!_spin_lock_try(lock));
 }
 

-extern mach_port_t thread_recycle_port;
+static mach_port_t thread_recycle_port = MACH_PORT_NULL;

 
 /* These are used to keep track of a semaphore pool shared by mutexes and condition
 ** variables.
 
 /* These are used to keep track of a semaphore pool shared by mutexes and condition
 ** variables.


@@ -163,24 +211,29 @@ static int pthread_concurrency;
 
 static OSSpinLock __workqueue_list_lock = OS_SPINLOCK_INIT;
 
 
 static OSSpinLock __workqueue_list_lock = OS_SPINLOCK_INIT;
 

-static void _pthread_exit(pthread_t self, void *value_ptr);
+static void _pthread_exit(pthread_t self, void *value_ptr) __dead2;

 static void _pthread_setcancelstate_exit(pthread_t self, void  *value_ptr, int conforming);
 static pthread_attr_t _pthread_attr_default = {0};
 static void _pthread_workq_init(pthread_workqueue_t wq, const pthread_workqueue_attr_t * attr);
 static int kernel_workq_setup = 0;
 static volatile int32_t kernel_workq_count = 0;
 static void _pthread_setcancelstate_exit(pthread_t self, void  *value_ptr, int conforming);
 static pthread_attr_t _pthread_attr_default = {0};
 static void _pthread_workq_init(pthread_workqueue_t wq, const pthread_workqueue_attr_t * attr);
 static int kernel_workq_setup = 0;
 static volatile int32_t kernel_workq_count = 0;

-static volatile unsigned int user_workq_count = 0; 
+static volatile unsigned int user_workq_count = 0;     /* number of outstanding workqueues */
+static volatile unsigned int user_workitem_count = 0;  /* number of outstanding workitems */

 #define KERNEL_WORKQ_ELEM_MAX          64              /* Max number of elements in the kerrel */
 static int wqreadyprio = 0;    /* current highest prio queue ready with items */
 
 #define KERNEL_WORKQ_ELEM_MAX          64              /* Max number of elements in the kerrel */
 static int wqreadyprio = 0;    /* current highest prio queue ready with items */
 

-static int __pthread_workqueue_affinity = 1;   /* 0 means no affinity */

 __private_extern__ struct __pthread_workitem_pool __pthread_workitem_pool_head = TAILQ_HEAD_INITIALIZER(__pthread_workitem_pool_head);
 __private_extern__ struct __pthread_workqueue_pool __pthread_workqueue_pool_head = TAILQ_HEAD_INITIALIZER(__pthread_workqueue_pool_head);
 
 __private_extern__ struct __pthread_workitem_pool __pthread_workitem_pool_head = TAILQ_HEAD_INITIALIZER(__pthread_workitem_pool_head);
 __private_extern__ struct __pthread_workqueue_pool __pthread_workqueue_pool_head = TAILQ_HEAD_INITIALIZER(__pthread_workqueue_pool_head);
 

+static struct _pthread_workitem *  __workqueue_pool_ptr;
+static size_t __workqueue_pool_size = 0;
+static int __workqueue_nitems = 0;
+

 struct _pthread_workqueue_head __pthread_workq0_head;
 struct _pthread_workqueue_head __pthread_workq1_head;
 struct _pthread_workqueue_head __pthread_workq2_head;
 struct _pthread_workqueue_head __pthread_workq0_head;
 struct _pthread_workqueue_head __pthread_workq1_head;
 struct _pthread_workqueue_head __pthread_workq2_head;

-pthread_workqueue_head_t __pthread_wq_head_tbl[WQ_NUM_PRIO_QS] = {&__pthread_workq0_head, &__pthread_workq1_head, &__pthread_workq2_head};
+struct _pthread_workqueue_head __pthread_workq3_head;
+pthread_workqueue_head_t __pthread_wq_head_tbl[WORKQ_NUM_PRIOQUEUE] = {&__pthread_workq0_head, &__pthread_workq1_head, &__pthread_workq2_head, &__pthread_workq3_head};

 
 static void workqueue_list_lock(void);
 static void workqueue_list_unlock(void);
 
 static void workqueue_list_lock(void);
 static void workqueue_list_unlock(void);


@@ -193,10 +246,12 @@ extern void start_wqthread(pthread_t self, mach_port_t kport, void * stackaddr,
 extern void thread_start(pthread_t self, mach_port_t kport, void *(*fun)(void *), void * funarg, size_t stacksize, unsigned int flags);
 static pthread_workitem_t alloc_workitem(void);
 static void free_workitem(pthread_workitem_t);
 extern void thread_start(pthread_t self, mach_port_t kport, void *(*fun)(void *), void * funarg, size_t stacksize, unsigned int flags);
 static pthread_workitem_t alloc_workitem(void);
 static void free_workitem(pthread_workitem_t);

+static void grow_workitem(void);

 static pthread_workqueue_t alloc_workqueue(void);
 static void free_workqueue(pthread_workqueue_t);
 static int _pthread_work_internal_init(void);
 static void workqueue_exit(pthread_t self, pthread_workqueue_t workq, pthread_workitem_t item);
 static pthread_workqueue_t alloc_workqueue(void);
 static void free_workqueue(pthread_workqueue_t);
 static int _pthread_work_internal_init(void);
 static void workqueue_exit(pthread_t self, pthread_workqueue_t workq, pthread_workitem_t item);

+void _pthread_fork_child_postinit();

 
 void pthread_workqueue_atfork_prepare(void);
 void pthread_workqueue_atfork_parent(void);
 
 void pthread_workqueue_atfork_prepare(void);
 void pthread_workqueue_atfork_parent(void);


@@ -211,6 +266,17 @@ extern void dispatch_atfork_child(void);
 #define WQOPS_QUEUE_REMOVE 2
 #define WQOPS_THREAD_RETURN 4
 #define WQOPS_THREAD_SETCONC  8
 #define WQOPS_QUEUE_REMOVE 2
 #define WQOPS_THREAD_RETURN 4
 #define WQOPS_THREAD_SETCONC  8

+#define WQOPS_QUEUE_NEWSPISUPP  0x10    /* this is to check for newer SPI support */
+#define WQOPS_QUEUE_REQTHREADS  0x20   /* request number of threads of a prio */
+
+/* flag values for reuse field in the libc side _pthread_wqthread */
+#define WQ_FLAG_THREAD_PRIOMASK                0x0000ffff
+#define WQ_FLAG_THREAD_OVERCOMMIT      0x00010000      /* thread is with overcommit prio */
+#define WQ_FLAG_THREAD_REUSE           0x00020000      /* thread is being reused */
+#define WQ_FLAG_THREAD_NEWSPI          0x00040000      /* the call is with new SPIs */
+
+
+#define WORKQUEUE_OVERCOMMIT 0x10000   /* the work_kernreturn() for overcommit in prio field */

 
 /*
  * Flags filed passed to bsdthread_create and back in pthread_start 
 
 /*
  * Flags filed passed to bsdthread_create and back in pthread_start 


@@ -234,7 +300,7 @@ void _pthread_wqthread(pthread_t self, mach_port_t kport, void * stackaddr, pthr
 
 static int pthread_setschedparam_internal(pthread_t, mach_port_t, int, const struct sched_param *);
 extern pthread_t __bsdthread_create(void *(*func)(void *), void * func_arg, void * stack, pthread_t  thread, unsigned int flags);
 
 static int pthread_setschedparam_internal(pthread_t, mach_port_t, int, const struct sched_param *);
 extern pthread_t __bsdthread_create(void *(*func)(void *), void * func_arg, void * stack, pthread_t  thread, unsigned int flags);

-extern int __bsdthread_register(void (*)(pthread_t, mach_port_t, void *(*)(void *), void *, size_t, unsigned int), void (*)(pthread_t, mach_port_t, void *, pthread_workitem_t, int), int,void (*)(pthread_t, mach_port_t, void *(*)(void *), void *, size_t, unsigned int), void (*)(pthread_t, mach_port_t, void *, pthread_workitem_t, int),__uint64_t);
+extern int __bsdthread_register(void (*)(pthread_t, mach_port_t, void *(*)(void *), void *, size_t, unsigned int), void (*)(pthread_t, mach_port_t, void *, pthread_workitem_t, int), int,void (*)(pthread_t, mach_port_t, void *(*)(void *), void *, size_t, unsigned int), int32_t *,__uint64_t);

 extern int __bsdthread_terminate(void * freeaddr, size_t freesize, mach_port_t kport, mach_port_t joinsem);
 extern __uint64_t __thread_selfid( void );
 extern int __pthread_canceled(int);
 extern int __bsdthread_terminate(void * freeaddr, size_t freesize, mach_port_t kport, mach_port_t joinsem);
 extern __uint64_t __thread_selfid( void );
 extern int __pthread_canceled(int);


@@ -243,7 +309,6 @@ extern int __pthread_kill(mach_port_t, int);
 extern int __pthread_markcancel(int);
 extern int __workq_open(void);
 
 extern int __pthread_markcancel(int);
 extern int __workq_open(void);
 

-#define WORKQUEUE_OVERCOMMIT 0x10000

 
 extern int __workq_kernreturn(int, pthread_workitem_t, int, int);
 
 
 extern int __workq_kernreturn(int, pthread_workitem_t, int, int);
 


@@ -251,7 +316,7 @@ extern int __workq_kernreturn(int, pthread_workitem_t, int, int);
 static const vm_address_t PTHREAD_STACK_HINT = 0xF0000000;
 #elif defined(__i386__) || defined(__x86_64__)
 static const vm_address_t PTHREAD_STACK_HINT = 0xB0000000;
 static const vm_address_t PTHREAD_STACK_HINT = 0xF0000000;
 #elif defined(__i386__) || defined(__x86_64__)
 static const vm_address_t PTHREAD_STACK_HINT = 0xB0000000;

-#elif defined(__arm__)
+#elif defined(__arm__)  

 static const vm_address_t PTHREAD_STACK_HINT = 0x30000000;
 #else
 #error Need to define a stack address hint for this architecture
 static const vm_address_t PTHREAD_STACK_HINT = 0x30000000;
 #else
 #error Need to define a stack address hint for this architecture


@@ -365,7 +430,6 @@ _pthread_free_pthread_onstack(pthread_t t, int freestruct, int termthread)
        kern_return_t res = 0;
        vm_address_t freeaddr;
        size_t freesize;
        kern_return_t res = 0;
        vm_address_t freeaddr;
        size_t freesize;

-       task_t self = mach_task_self();

        int thread_count;
        mach_port_t kport;
        semaphore_t joinsem = SEMAPHORE_NULL;
        int thread_count;
        mach_port_t kport;
        semaphore_t joinsem = SEMAPHORE_NULL;


@@ -820,13 +884,16 @@ static void
 _pthread_body(pthread_t self)
 {
     _pthread_set_self(self);
 _pthread_body(pthread_t self)
 {
     _pthread_set_self(self);

+#if defined(__i386__) || defined(__x86_64__) || defined(__arm__) 
+       if( (self->thread_id = __thread_selfid()) == (__uint64_t)-1)
+               printf("Failed to set thread_id in _pthread_body\n");
+#endif

     _pthread_exit(self, (self->fun)(self->arg));
 }
 
 void
 _pthread_start(pthread_t self, mach_port_t kport, void *(*fun)(void *), void * funarg, size_t stacksize, unsigned int pflags)
 {
     _pthread_exit(self, (self->fun)(self->arg));
 }
 
 void
 _pthread_start(pthread_t self, mach_port_t kport, void *(*fun)(void *), void * funarg, size_t stacksize, unsigned int pflags)
 {

-       int ret;

 #if WQ_DEBUG
        pthread_t pself;
 #endif
 #if WQ_DEBUG
        pthread_t pself;
 #endif


@@ -836,7 +903,7 @@ _pthread_start(pthread_t self, mach_port_t kport, void *(*fun)(void *), void * f
        if ((pflags & PTHREAD_START_CUSTOM) == 0) {
                stackaddr = (char *)self;
                _pthread_struct_init(self, attrs, stackaddr,  stacksize, 1, 1);
        if ((pflags & PTHREAD_START_CUSTOM) == 0) {
                stackaddr = (char *)self;
                _pthread_struct_init(self, attrs, stackaddr,  stacksize, 1, 1);

-#if defined(__i386__) || defined(__x86_64__) || defined(__arm__)
+#if defined(__i386__) || defined(__x86_64__) || defined(__arm__) 

                _pthread_set_self(self);
 #endif
                LOCK(_pthread_list_lock);
                _pthread_set_self(self);
 #endif
                LOCK(_pthread_list_lock);


@@ -850,7 +917,7 @@ _pthread_start(pthread_t self, mach_port_t kport, void *(*fun)(void *), void * f
                        self->detached |= PTHREAD_CREATE_DETACHED;
                }
        }  else { 
                        self->detached |= PTHREAD_CREATE_DETACHED;
                }
        }  else { 

-#if defined(__i386__) || defined(__x86_64__) || defined(__arm__)
+#if defined(__i386__) || defined(__x86_64__) || defined(__arm__) 

                _pthread_set_self(self);
 #endif
                LOCK(_pthread_list_lock);
                _pthread_set_self(self);
 #endif
                LOCK(_pthread_list_lock);


@@ -870,7 +937,7 @@ _pthread_start(pthread_t self, mach_port_t kport, void *(*fun)(void *), void * f
        self->childrun = 1;
        UNLOCK(_pthread_list_lock);
 
        self->childrun = 1;
        UNLOCK(_pthread_list_lock);
 

-#if defined(__i386__) || defined(__x86_64__) || defined(__arm__)
+#if defined(__i386__) || defined(__x86_64__) || defined(__arm__) 

        if( (self->thread_id = __thread_selfid()) == (__uint64_t)-1)
                printf("Failed to set thread_id in pthread_start\n");
 #endif
        if( (self->thread_id = __thread_selfid()) == (__uint64_t)-1)
                printf("Failed to set thread_id in pthread_start\n");
 #endif


@@ -916,7 +983,7 @@ _pthread_create(pthread_t t,
                t->policy = attrs->policy;
                t->param = attrs->param;
                t->freeStackOnExit = attrs->freeStackOnExit;
                t->policy = attrs->policy;
                t->param = attrs->param;
                t->freeStackOnExit = attrs->freeStackOnExit;

-               t->mutexes = (struct _pthread_mutex *)NULL;
+               t->cancel_error = 0;

                t->sig = _PTHREAD_SIG;
                 t->reply_port = MACH_PORT_NULL;
                 t->cthread_self = NULL;
                t->sig = _PTHREAD_SIG;
                 t->reply_port = MACH_PORT_NULL;
                 t->cthread_self = NULL;


@@ -936,7 +1003,7 @@ _pthread_create(pthread_t t,
 void
 _pthread_struct_init(pthread_t t, const pthread_attr_t *attrs, void * stack, size_t stacksize, int kernalloc, int nozero)
 {
 void
 _pthread_struct_init(pthread_t t, const pthread_attr_t *attrs, void * stack, size_t stacksize, int kernalloc, int nozero)
 {

-       mach_vm_offset_t stackaddr = (mach_vm_offset_t)(long)stack;
+       mach_vm_offset_t stackaddr = (mach_vm_offset_t)(uintptr_t)stack;

 
                if (nozero == 0) {
                        memset(t, 0, sizeof(*t));
 
                if (nozero == 0) {
                        memset(t, 0, sizeof(*t));


@@ -946,13 +1013,13 @@ _pthread_struct_init(pthread_t t, const pthread_attr_t *attrs, void * stack, siz
                t->schedset = attrs->schedset;
                t->tsd[0] = t;
                if (kernalloc != 0) {
                t->schedset = attrs->schedset;
                t->tsd[0] = t;
                if (kernalloc != 0) {

-                       stackaddr = (mach_vm_offset_t)(long)t;
+                       stackaddr = (mach_vm_offset_t)(uintptr_t)t;

 
                        /* if allocated from kernel set values appropriately */
                        t->stacksize = stacksize;
 
                        /* if allocated from kernel set values appropriately */
                        t->stacksize = stacksize;

-                               t->stackaddr = (void *)(long)stackaddr;
+                               t->stackaddr = (void *)(uintptr_t)stackaddr;

                        t->freeStackOnExit = 1;
                        t->freeStackOnExit = 1;

-                       t->freeaddr = (void *)(long)(stackaddr - stacksize - vm_page_size);
+                       t->freeaddr = (void *)(uintptr_t)(stackaddr - stacksize - vm_page_size);

                        t->freesize = pthreadsize + stacksize + vm_page_size;
                } else {
                        t->stacksize = attrs->stacksize;
                        t->freesize = pthreadsize + stacksize + vm_page_size;
                } else {
                        t->stacksize = attrs->stacksize;


@@ -963,7 +1030,7 @@ _pthread_struct_init(pthread_t t, const pthread_attr_t *attrs, void * stack, siz
                t->inherit = attrs->inherit;
                t->policy = attrs->policy;
                t->param = attrs->param;
                t->inherit = attrs->inherit;
                t->policy = attrs->policy;
                t->param = attrs->param;

-               t->mutexes = (struct _pthread_mutex *)NULL;
+               t->cancel_error = 0;

                t->sig = _PTHREAD_SIG;
                t->reply_port = MACH_PORT_NULL;
                t->cthread_self = NULL;
                t->sig = _PTHREAD_SIG;
                t->reply_port = MACH_PORT_NULL;
                t->cthread_self = NULL;


@@ -979,13 +1046,6 @@ _pthread_struct_init(pthread_t t, const pthread_attr_t *attrs, void * stack, siz
                t->max_tsd_key = 0;
 }
 
                t->max_tsd_key = 0;
 }
 

-static void
-_pthread_tsd_reinit(pthread_t t)
-{
-       bzero(&t->tsd[1], (_INTERNAL_POSIX_THREAD_KEYS_END-1) * sizeof(void *));
-}
-
-

 /* Need to deprecate this in future */
 int
 _pthread_is_threaded(void)
 /* Need to deprecate this in future */
 int
 _pthread_is_threaded(void)


@@ -1047,12 +1107,8 @@ pthread_t pthread_from_mach_thread_np(mach_port_t kernel_thread)
 size_t
 pthread_get_stacksize_np(pthread_t t)
 {
 size_t
 pthread_get_stacksize_np(pthread_t t)
 {

-       int ret,nestingDepth=0;
+       int ret;

        size_t size = 0;
        size_t size = 0;

-       vm_address_t address=0;
-        vm_size_t region_size=0;
-       struct vm_region_submap_info_64 info;
-       mach_msg_type_number_t  count;

 
        if (t == NULL)
                return(ESRCH);
 
        if (t == NULL)
                return(ESRCH);


@@ -1084,7 +1140,7 @@ pthread_get_stackaddr_np(pthread_t t)
        void * addr = NULL;
 
        if (t == NULL)
        void * addr = NULL;
 
        if (t == NULL)

-               return((void *)(long)ESRCH);
+               return((void *)(uintptr_t)ESRCH);

        
        if(t == pthread_self() || t == &_thread) //since the main thread will not get deallocated from underneath us
                return t->stackaddr;
        
        if(t == pthread_self() || t == &_thread) //since the main thread will not get deallocated from underneath us
                return t->stackaddr;


@@ -1093,7 +1149,7 @@ pthread_get_stackaddr_np(pthread_t t)
 
        if ((ret = _pthread_find_thread(t)) != 0) {
                UNLOCK(_pthread_list_lock);
 
        if ((ret = _pthread_find_thread(t)) != 0) {
                UNLOCK(_pthread_list_lock);

-               return((void *)(long)ret);
+               return((void *)(uintptr_t)ret);

        }
        addr = t->stackaddr;
        UNLOCK(_pthread_list_lock);
        }
        addr = t->stackaddr;
        UNLOCK(_pthread_list_lock);


@@ -1118,7 +1174,7 @@ pthread_main_np(void)
 }
 
 
 }
 
 

-#if defined(__i386__) || defined(__x86_64__) || defined(__arm__)
+#if defined(__i386__) || defined(__x86_64__) || defined(__arm__) 

 /* if we are passed in a pthread_t that is NULL, then we return
    the current thread's thread_id. So folks don't have to call
    pthread_self, in addition to us doing it, if they just want 
 /* if we are passed in a pthread_t that is NULL, then we return
    the current thread's thread_id. So folks don't have to call
    pthread_self, in addition to us doing it, if they just want 


@@ -1170,15 +1226,22 @@ pthread_getname_np(pthread_t thread, char *threadname, size_t len)
 int
 pthread_setname_np(const char *threadname)
 {
 int
 pthread_setname_np(const char *threadname)
 {

-       int rval;
-       size_t len;
+       int rval = 0;
+       int len = 0;
+       pthread_t current_thread = pthread_self();
+
+       if (threadname != NULL)
+               len = strlen(threadname);
+
+       /* protytype is in pthread_internals.h */
+       rval = proc_setthreadname((void *)threadname, len);
+       if (rval == 0) {
+               if (threadname != NULL) {
+                       strlcpy(current_thread->pthread_name, threadname, MAXTHREADNAMESIZE);
+               } else {
+                       memset(current_thread->pthread_name, 0 , MAXTHREADNAMESIZE);
+               }

 
 

-       rval = 0;
-       len = strlen(threadname);
-       rval = sysctlbyname("kern.threadname", NULL, 0, threadname, len);
-       if(rval == 0)
-       {
-               strlcpy((pthread_self())->pthread_name, threadname, len+1);     

        }
        return rval;
 
        }
        return rval;
 


@@ -1309,7 +1372,7 @@ _new_pthread_create_suspended(pthread_t *thread,
                        __kdebug_trace(0x9000008, t, 0, 0, 1, 0);
 #endif
                        if(t->freeStackOnExit)
                        __kdebug_trace(0x9000008, t, 0, 0, 1, 0);
 #endif
                        if(t->freeStackOnExit)

-                                vm_deallocate(self, (mach_vm_address_t)(long)t, pthreadsize);
+                                vm_deallocate(self, (mach_vm_address_t)(uintptr_t)t, pthreadsize);

                         else
                                free(t);
                        } else if (t->childrun == 0) {
                         else
                                free(t);
                        } else if (t->childrun == 0) {


@@ -1345,7 +1408,7 @@ _new_pthread_create_suspended(pthread_t *thread,
 #if PTH_TRACE
                        __kdebug_trace(0x9000008, t, pthreadsize, 0, 2, 0);
 #endif
 #if PTH_TRACE
                        __kdebug_trace(0x9000008, t, pthreadsize, 0, 2, 0);
 #endif

-                               vm_deallocate(self, (mach_vm_address_t)(long)t, pthreadsize);
+                               vm_deallocate(self, (mach_vm_address_t)(uintptr_t)t, pthreadsize);

                        } else if (t->childrun == 0) {
                                TAILQ_INSERT_TAIL(&__pthread_head, t, plist);
                                _pthread_count++;
                        } else if (t->childrun == 0) {
                                TAILQ_INSERT_TAIL(&__pthread_head, t, plist);
                                _pthread_count++;


@@ -1467,9 +1530,10 @@ pthread_detach(pthread_t thread)
        int newstyle = 0;
        int ret;
 
        int newstyle = 0;
        int ret;
 

-       if ((ret = _pthread_lookup_thread(thread, NULL, 1)) != 0)
+       if ((ret = _pthread_lookup_thread(thread, NULL, 1)) != 0) {

                return (ret); /* Not a valid thread */
                return (ret); /* Not a valid thread */

-               
+       }
+

        LOCK(thread->lock);
        newstyle = thread->newstyle;
        if (thread->detached & PTHREAD_CREATE_JOINABLE)
        LOCK(thread->lock);
        newstyle = thread->newstyle;
        if (thread->detached & PTHREAD_CREATE_JOINABLE)


@@ -1507,7 +1571,7 @@ pthread_detach(pthread_t thread)
 }
 
 
 }
 
 

-/* 
+/*

  * pthread_kill call to system call
  */
 int   
  * pthread_kill call to system call
  */
 int   


@@ -1921,16 +1985,21 @@ pthread_equal(pthread_t t1,
        return (t1 == t2);
 }
 
        return (t1 == t2);
 }
 

-__private_extern__ void
+// Force LLVM not to optimise this to a call to __pthread_set_self, if it does
+// then _pthread_set_self won't be bound when secondary threads try and start up.
+void __attribute__((noinline))

 _pthread_set_self(pthread_t p)
 {
 _pthread_set_self(pthread_t p)
 {

-       extern void __pthread_set_self(pthread_t);
+       extern void __pthread_set_self(void *);
+

         if (p == 0) {
         if (p == 0) {

-                bzero(&_thread, sizeof(struct _pthread));
+                if (_thread.tsd[0] != 0) {
+                        bzero(&_thread, sizeof(struct _pthread));
+                }

                 p = &_thread;
         }
         p->tsd[0] = p;
                 p = &_thread;
         }
         p->tsd[0] = p;

-       __pthread_set_self(p);
+       __pthread_set_self(&p->tsd[0]);

 }
 
 void 
 }
 
 void 


@@ -2022,7 +2091,7 @@ pthread_setconcurrency(int new_level)
 /*
  * Perform package initialization - called automatically when application starts
  */
 /*
  * Perform package initialization - called automatically when application starts
  */

-__private_extern__ int
+int

 pthread_init(void)
 {
        pthread_attr_t *attrs;
 pthread_init(void)
 {
        pthread_attr_t *attrs;


@@ -2082,6 +2151,7 @@ pthread_init(void)
        workq_targetconc[WORKQ_HIGH_PRIOQUEUE] = ncpus;
        workq_targetconc[WORKQ_DEFAULT_PRIOQUEUE] = ncpus;
        workq_targetconc[WORKQ_LOW_PRIOQUEUE] =  ncpus;
        workq_targetconc[WORKQ_HIGH_PRIOQUEUE] = ncpus;
        workq_targetconc[WORKQ_DEFAULT_PRIOQUEUE] = ncpus;
        workq_targetconc[WORKQ_LOW_PRIOQUEUE] =  ncpus;

+       workq_targetconc[WORKQ_BG_PRIOQUEUE] =  ncpus;

 
        mach_port_deallocate(mach_task_self(), host);
     
 
        mach_port_deallocate(mach_task_self(), host);
     


@@ -2104,21 +2174,22 @@ pthread_init(void)
                /* We ignore the return result here. The ObjC runtime will just have to deal. */
 }
 #endif
                /* We ignore the return result here. The ObjC runtime will just have to deal. */
 }
 #endif

-
+       //added so that thread_recycle_port is initialized on new launch.
+       _pthread_fork_child_postinit();

        mig_init(1);            /* enable multi-threaded mig interfaces */
        if (__oldstyle == 0) {
        mig_init(1);            /* enable multi-threaded mig interfaces */
        if (__oldstyle == 0) {

-#if defined(__i386__) || defined(__x86_64__) || defined(__arm__)
-               __bsdthread_register(thread_start, start_wqthread, round_page(sizeof(struct _pthread)), _pthread_start, &workq_targetconc[0], (__uint64_t)(&thread->tsd[__PTK_LIBDISPATCH_KEY0]) - (__uint64_t)thread);
+#if defined(__i386__) || defined(__x86_64__) || defined(__arm__) 
+               __bsdthread_register(thread_start, start_wqthread, round_page(sizeof(struct _pthread)), _pthread_start, &workq_targetconc[0], (uintptr_t)(&thread->tsd[__PTK_LIBDISPATCH_KEY0]) - (uintptr_t)(&thread->tsd[0]));

 #else
 #else

-               __bsdthread_register(_pthread_start, _pthread_wqthread, round_page(sizeof(struct _pthread)), NULL, &workq_targetconc[0], (__uint64_t)&thread->tsd[__PTK_LIBDISPATCH_KEY0] - (__uint64_t)thread);
+               __bsdthread_register(_pthread_start, _pthread_wqthread, round_page(sizeof(struct _pthread)), NULL, &workq_targetconc[0], (uintptr_t)&thread->tsd[__PTK_LIBDISPATCH_KEY0] - (uintptr_t)thread);

 #endif
        }
 
 #endif
        }
 

-#if defined(__i386__) || defined(__x86_64__) || defined(__arm__)
+#if defined(__i386__) || defined(__x86_64__) || defined(__arm__) 

        if( (thread->thread_id = __thread_selfid()) == (__uint64_t)-1)
                printf("Failed to set thread_id in pthread_init\n");
        if( (thread->thread_id = __thread_selfid()) == (__uint64_t)-1)
                printf("Failed to set thread_id in pthread_init\n");

-       return 0;

 #endif
 #endif

+       return 0;

 }
 
 int sched_yield(void)
 }
 
 int sched_yield(void)


@@ -2128,8 +2199,12 @@ int sched_yield(void)
 }
 
 /* This used to be the "magic" that gets the initialization routine called when the application starts */
 }
 
 /* This used to be the "magic" that gets the initialization routine called when the application starts */

-static int _do_nothing(void) { return 0; }
-int (*_cthread_init_routine)(void) = _do_nothing;
+/*
+ * (These has been moved to setenv.c, so we can use it to fix a less than 10.5
+ * crt1.o issue)
+ * static int _do_nothing(void) { return 0; }
+ * int (*_cthread_init_routine)(void) = _do_nothing;
+ */

 
 /* Get a semaphore from the pool, growing it if necessary */
 
 
 /* Get a semaphore from the pool, growing it if necessary */
 


@@ -2180,12 +2255,21 @@ __private_extern__ void _pthread_fork_child(pthread_t p) {
        __kdebug_trace(0x900000c, p, 0, 0, 10, 0);
 #endif
        _pthread_count = 1;
        __kdebug_trace(0x900000c, p, 0, 0, 10, 0);
 #endif
        _pthread_count = 1;

-#if defined(__i386__) || defined(__x86_64__) || defined(__arm__)
+#if defined(__i386__) || defined(__x86_64__) || defined(__arm__) 

        if( (p->thread_id = __thread_selfid()) == (__uint64_t)-1)
                printf("Failed to set thread_id in pthread_fork_child\n");
 #endif
 }
 
        if( (p->thread_id = __thread_selfid()) == (__uint64_t)-1)
                printf("Failed to set thread_id in pthread_fork_child\n");
 #endif
 }
 

+void _pthread_fork_child_postinit() {
+       kern_return_t kr;
+
+       kr = mach_port_allocate(mach_task_self(), MACH_PORT_RIGHT_RECEIVE, &thread_recycle_port);
+       if (kr != KERN_SUCCESS) {
+               abort();
+       }
+}
+

 /*
  * Query/update the cancelability 'state' of a thread
  */
 /*
  * Query/update the cancelability 'state' of a thread
  */


@@ -2239,7 +2323,7 @@ int
 _pthread_join_cleanup(pthread_t thread, void ** value_ptr, int conforming)
 {
        kern_return_t res;
 _pthread_join_cleanup(pthread_t thread, void ** value_ptr, int conforming)
 {
        kern_return_t res;

-       int detached = 0, ret;
+       int ret;

 
 #if PTH_TRACE
        __kdebug_trace(0x9000028, thread, 0, 0, 1, 0);
 
 #if PTH_TRACE
        __kdebug_trace(0x9000028, thread, 0, 0, 1, 0);


@@ -2288,7 +2372,7 @@ _pthread_join_cleanup(pthread_t thread, void ** value_ptr, int conforming)
 #if PTH_TRACE
        __kdebug_trace(0x9000028, thread, 0, 0, 2, 0);
 #endif
 #if PTH_TRACE
        __kdebug_trace(0x9000028, thread, 0, 0, 2, 0);
 #endif

-               vm_deallocate(mach_task_self(), (mach_vm_address_t)(long)thread, pthreadsize);
+               vm_deallocate(mach_task_self(), (mach_vm_address_t)(uintptr_t)thread, pthreadsize);

        } else {
                thread->sig = _PTHREAD_NO_SIG;
 #if PTH_TRACE
        } else {
                thread->sig = _PTHREAD_NO_SIG;
 #if PTH_TRACE


@@ -2390,6 +2474,7 @@ int error = 0;
                        case WORKQ_HIGH_PRIOQUEUE:
                        case WORKQ_DEFAULT_PRIOQUEUE:
                        case WORKQ_LOW_PRIOQUEUE:
                        case WORKQ_HIGH_PRIOQUEUE:
                        case WORKQ_DEFAULT_PRIOQUEUE:
                        case WORKQ_LOW_PRIOQUEUE:

+                       case WORKQ_BG_PRIOQUEUE:

                                attr->queueprio = qprio;
                                break;
                        default:
                                attr->queueprio = qprio;
                                break;
                        default:


@@ -2445,6 +2530,10 @@ pthread_workqueue_init_np()
 {
        int ret;
 
 {
        int ret;
 

+       if (__workqueue_newspis != 0)
+               return(EPERM);
+       __workqueue_oldspis = 1;        
+

        workqueue_list_lock();
        ret =_pthread_work_internal_init();
        workqueue_list_unlock();
        workqueue_list_lock();
        ret =_pthread_work_internal_init();
        workqueue_list_unlock();


@@ -2457,6 +2546,9 @@ pthread_workqueue_requestconcurrency_np(int queue, int request_concurrency)
 {
        int error = 0;
 
 {
        int error = 0;
 

+       if (__workqueue_newspis != 0)
+               return(EPERM);
+

        if (queue < 0 || queue > WORKQ_NUM_PRIOQUEUE)
                return(EINVAL);
 
        if (queue < 0 || queue > WORKQ_NUM_PRIOQUEUE)
                return(EINVAL);
 


@@ -2490,13 +2582,33 @@ pthread_workqueue_atfork_parent(void)
 void
 pthread_workqueue_atfork_child(void)
 {
 void
 pthread_workqueue_atfork_child(void)
 {

-#if defined(__i386__) || defined(__x86_64__) || defined(__arm__)
+#if defined(__i386__) || defined(__x86_64__) || defined(__arm__) 
+       pthread_t self = pthread_self();
+
+       __workqueue_list_lock = OS_SPINLOCK_INIT;
+
+       /* already using new spis? */
+       if (__workqueue_newspis != 0) {
+               /* prepare the kernel for workq action */
+#if defined(__i386__) || defined(__x86_64__) || defined(__arm__) 
+               __bsdthread_register(thread_start, start_wqthread, round_page(sizeof(struct _pthread)), _pthread_start, &workq_targetconc[0], (uintptr_t)(&self->tsd[__PTK_LIBDISPATCH_KEY0]) - (uintptr_t)(&self->tsd[0]));
+#else
+               __bsdthread_register(_pthread_start, _pthread_wqthread, round_page(sizeof(struct _pthread)),NULL,NULL,0);
+#endif
+               (void)__workq_open();
+               kernel_workq_setup = 1;
+               return;
+       }
+
+       /* not using old spis either? */
+       if (__workqueue_oldspis == 0)
+               return;
+

        /* 
         * NOTE:  workq additions here  
         * are for i386,x86_64 only as
         * ppc and arm do not support it
         */
        /* 
         * NOTE:  workq additions here  
         * are for i386,x86_64 only as
         * ppc and arm do not support it
         */

-       __workqueue_list_lock = OS_SPINLOCK_INIT;

        if (kernel_workq_setup != 0){
           kernel_workq_setup = 0;
           _pthread_work_internal_init();
        if (kernel_workq_setup != 0){
           kernel_workq_setup = 0;
           _pthread_work_internal_init();


@@ -2510,31 +2622,47 @@ _pthread_work_internal_init(void)
 {
        int i, error;
        pthread_workqueue_head_t headp;
 {
        int i, error;
        pthread_workqueue_head_t headp;

-       pthread_workitem_t witemp;

        pthread_workqueue_t wq;
        pthread_workqueue_t wq;

+#if defined(__i386__) || defined(__x86_64__) || defined(__arm__) 
+       pthread_t self = pthread_self();
+#endif

 
        if (kernel_workq_setup == 0) {
 
        if (kernel_workq_setup == 0) {

-#if defined(__i386__) || defined(__x86_64__) || defined(__arm__)
-               __bsdthread_register(thread_start, start_wqthread, round_page(sizeof(struct _pthread)),NULL,NULL, NULL);
+#if defined(__i386__) || defined(__x86_64__) || defined(__arm__) 
+               __bsdthread_register(thread_start, start_wqthread, round_page(sizeof(struct _pthread)), _pthread_start, &workq_targetconc[0], (uintptr_t)(&self->tsd[__PTK_LIBDISPATCH_KEY0]) - (uintptr_t)(&self->tsd[0]));

 #else
 #else

-               __bsdthread_register(_pthread_start, _pthread_wqthread, round_page(sizeof(struct _pthread)),NULL,NULL, NULL);
+               __bsdthread_register(_pthread_start, _pthread_wqthread, round_page(sizeof(struct _pthread)),NULL,NULL,0);

 #endif
 
                _pthread_wq_attr_default.queueprio = WORKQ_DEFAULT_PRIOQUEUE;
                _pthread_wq_attr_default.sig = PTHREAD_WORKQUEUE_ATTR_SIG;
 
 #endif
 
                _pthread_wq_attr_default.queueprio = WORKQ_DEFAULT_PRIOQUEUE;
                _pthread_wq_attr_default.sig = PTHREAD_WORKQUEUE_ATTR_SIG;
 

-               for( i = 0; i< WQ_NUM_PRIO_QS; i++) {
+               for( i = 0; i< WORKQ_NUM_PRIOQUEUE; i++) {

                        headp = __pthread_wq_head_tbl[i];
                        TAILQ_INIT(&headp->wqhead);
                        headp->next_workq = 0;
                }
 
                        headp = __pthread_wq_head_tbl[i];
                        TAILQ_INIT(&headp->wqhead);
                        headp->next_workq = 0;
                }
 

-               /* create work item and workqueue pools */
-               witemp = (struct _pthread_workitem *)malloc(sizeof(struct _pthread_workitem) * WORKITEM_POOL_SIZE);
-               bzero(witemp, (sizeof(struct _pthread_workitem) * WORKITEM_POOL_SIZE));
-               for (i = 0; i < WORKITEM_POOL_SIZE; i++) {
-                       TAILQ_INSERT_TAIL(&__pthread_workitem_pool_head, &witemp[i], item_entry);
-               }
+               __workqueue_pool_ptr = NULL;
+               __workqueue_pool_size = round_page(sizeof(struct _pthread_workitem) * WORKITEM_POOL_SIZE);
+
+               __workqueue_pool_ptr = (struct _pthread_workitem *)mmap(NULL, __workqueue_pool_size,
+                       PROT_READ|PROT_WRITE,
+                       MAP_ANON | MAP_PRIVATE,
+                       0,
+                       0);
+
+               if (__workqueue_pool_ptr == MAP_FAILED) {
+                       /* Not expected to fail, if it does, always malloc for work items */
+                       __workqueue_nitems = WORKITEM_POOL_SIZE;
+                       __workqueue_pool_ptr = NULL;
+               } else 
+                       __workqueue_nitems = 0;
+
+               /* sets up the workitem pool */
+               grow_workitem();
+
+               /* since the size is less than a page, leaving this in malloc pool */

                wq = (struct _pthread_workqueue *)malloc(sizeof(struct _pthread_workqueue) * WORKQUEUE_POOL_SIZE);
                bzero(wq, (sizeof(struct _pthread_workqueue) * WORKQUEUE_POOL_SIZE));
                for (i = 0; i < WORKQUEUE_POOL_SIZE; i++) {
                wq = (struct _pthread_workqueue *)malloc(sizeof(struct _pthread_workqueue) * WORKQUEUE_POOL_SIZE);
                bzero(wq, (sizeof(struct _pthread_workqueue) * WORKQUEUE_POOL_SIZE));
                for (i = 0; i < WORKQUEUE_POOL_SIZE; i++) {


@@ -2544,7 +2672,9 @@ _pthread_work_internal_init(void)
                if (error = __workq_open()) {
                        TAILQ_INIT(&__pthread_workitem_pool_head);
                        TAILQ_INIT(&__pthread_workqueue_pool_head);
                if (error = __workq_open()) {
                        TAILQ_INIT(&__pthread_workitem_pool_head);
                        TAILQ_INIT(&__pthread_workqueue_pool_head);

-                       free(witemp);
+                       if (__workqueue_pool_ptr != NULL) {
+                               munmap((void *)__workqueue_pool_ptr, __workqueue_pool_size);
+                       }

                        free(wq);
                        return(ENOMEM);
                }
                        free(wq);
                        return(ENOMEM);
                }


@@ -2561,14 +2691,25 @@ alloc_workitem(void)
        pthread_workitem_t witem;
 
        if (TAILQ_EMPTY(&__pthread_workitem_pool_head)) {
        pthread_workitem_t witem;
 
        if (TAILQ_EMPTY(&__pthread_workitem_pool_head)) {

-               workqueue_list_unlock();
-               witem = malloc(sizeof(struct _pthread_workitem));
-               witem->gencount = 0;
-               workqueue_list_lock();
-       } else {
-               witem = TAILQ_FIRST(&__pthread_workitem_pool_head);
-               TAILQ_REMOVE(&__pthread_workitem_pool_head, witem, item_entry);
+               /* the chunk size is set so some multiple of it is pool size */
+               if (__workqueue_nitems < WORKITEM_POOL_SIZE) {
+                       grow_workitem();
+               } else {
+                       workqueue_list_unlock();
+                       witem = malloc(sizeof(struct _pthread_workitem));
+                       workqueue_list_lock();
+                       witem->fromcache = 0;
+                       goto out;
+               }

        }
        }

+       witem = TAILQ_FIRST(&__pthread_workitem_pool_head);
+       TAILQ_REMOVE(&__pthread_workitem_pool_head, witem, item_entry);
+       witem->fromcache = 1;
+out:
+       witem->flags = 0;
+       witem->item_entry.tqe_next = 0;
+       witem->item_entry.tqe_prev = 0;
+       user_workitem_count++;

        return(witem);
 }
 
        return(witem);
 }
 


@@ -2576,8 +2717,27 @@ alloc_workitem(void)
 static void
 free_workitem(pthread_workitem_t witem) 
 {
 static void
 free_workitem(pthread_workitem_t witem) 
 {

-       witem->gencount++;
-       TAILQ_INSERT_TAIL(&__pthread_workitem_pool_head, witem, item_entry);
+       user_workitem_count--;
+       witem->flags = 0;
+       if (witem->fromcache != 0)
+               TAILQ_INSERT_TAIL(&__pthread_workitem_pool_head, witem, item_entry);
+       else
+               free(witem);
+}
+
+static void
+grow_workitem(void)
+{
+       pthread_workitem_t witemp;
+       int i;
+
+        witemp = &__workqueue_pool_ptr[__workqueue_nitems];
+       bzero(witemp, (sizeof(struct _pthread_workitem) * WORKITEM_CHUNK_SIZE));
+       for (i = 0; i < WORKITEM_CHUNK_SIZE; i++) {
+               witemp[i].fromcache = 1;
+               TAILQ_INSERT_TAIL(&__pthread_workitem_pool_head, &witemp[i], item_entry);
+       }
+       __workqueue_nitems += WORKITEM_CHUNK_SIZE;

 }
 
 /* This routine is called with list lock held */
 }
 
 /* This routine is called with list lock held */


@@ -2655,7 +2815,7 @@ pick_nextworkqueue_droplock()
 loop:
        while (kernel_workq_count < KERNEL_WORKQ_ELEM_MAX) {
                found = 0;
 loop:
        while (kernel_workq_count < KERNEL_WORKQ_ELEM_MAX) {
                found = 0;

-               for (i = 0; i < WQ_NUM_PRIO_QS; i++)  {
+               for (i = 0; i < WORKQ_NUM_PRIOQUEUE; i++)  {

                        wqreadyprio = i;        /* because there is nothing else higher to run */
                        headp = __pthread_wq_head_tbl[i];
 
                        wqreadyprio = i;        /* because there is nothing else higher to run */
                        headp = __pthread_wq_head_tbl[i];
 


@@ -2776,7 +2936,6 @@ post_nextworkitem(pthread_workqueue_t workq)
 #if WQ_LISTTRACE
        __kdebug_trace(0x90080a8, workq, &workq->item_listhead, workq->item_listhead.tqh_first,  workq->item_listhead.tqh_last, 0);
 #endif
 #if WQ_LISTTRACE
        __kdebug_trace(0x90080a8, workq, &workq->item_listhead, workq->item_listhead.tqh_first,  workq->item_listhead.tqh_last, 0);
 #endif

-                               witem->flags = 0;

                                free_workitem(witem);
 #if WQ_TRACE
                        __kdebug_trace(0x9000064, 4, workq->barrier_count, 0, 0, 0);
                                free_workitem(witem);
 #if WQ_TRACE
                        __kdebug_trace(0x9000064, 4, workq->barrier_count, 0, 0, 0);


@@ -2805,7 +2964,6 @@ post_nextworkitem(pthread_workqueue_t workq)
                                __kdebug_trace(0x900006c, workq, workq->kq_count, 0, 0xff, 0);
 #endif
                                }
                                __kdebug_trace(0x900006c, workq, workq->kq_count, 0, 0xff, 0);
 #endif
                                }

-                               witem->flags = 0;

                                free_workitem(witem);
                                workq->flags |= PTHREAD_WORKQ_DESTROYED;
 #if WQ_TRACE
                                free_workitem(witem);
                                workq->flags |= PTHREAD_WORKQ_DESTROYED;
 #if WQ_TRACE


@@ -2852,7 +3010,7 @@ post_nextworkitem(pthread_workqueue_t workq)
                                workq->kq_count++;
                                witem->flags |= PTH_WQITEM_KERN_COUNT;
                        }
                                workq->kq_count++;
                                witem->flags |= PTH_WQITEM_KERN_COUNT;
                        }

-                       OSAtomicIncrement32(&kernel_workq_count);
+                       OSAtomicIncrement32Barrier(&kernel_workq_count);

                        workqueue_list_unlock();
                        
                        prio = workq->queueprio;
                        workqueue_list_unlock();
                        
                        prio = workq->queueprio;


@@ -2861,7 +3019,7 @@ post_nextworkitem(pthread_workqueue_t workq)
                        }
 
                        if (( error =__workq_kernreturn(WQOPS_QUEUE_ADD, witem, workq->affinity, prio)) == -1) {
                        }
 
                        if (( error =__workq_kernreturn(WQOPS_QUEUE_ADD, witem, workq->affinity, prio)) == -1) {

-                               OSAtomicDecrement32(&kernel_workq_count);
+                               OSAtomicDecrement32Barrier(&kernel_workq_count);

                                workqueue_list_lock();
 #if WQ_TRACE
                        __kdebug_trace(0x900007c, witem, workq, witem->func_arg, workq->kq_count, 0);
                                workqueue_list_lock();
 #if WQ_TRACE
                        __kdebug_trace(0x900007c, witem, workq, witem->func_arg, workq->kq_count, 0);


@@ -2897,10 +3055,25 @@ _pthread_wqthread(pthread_t self, mach_port_t kport, void * stackaddr, pthread_w
 #if WQ_DEBUG
        pthread_t pself;
 #endif
 #if WQ_DEBUG
        pthread_t pself;
 #endif

+       int thread_reuse = 0;
+       int thread_priority = 0;
+       int thread_newspi = 0;
+       int thread_options = 0;
+
+       if (reuse & WQ_FLAG_THREAD_NEWSPI) {
+               thread_reuse = reuse & WQ_FLAG_THREAD_REUSE;
+               if ((reuse & WQ_FLAG_THREAD_OVERCOMMIT) != 0)
+                       thread_options = WORKQ_ADDTHREADS_OPTION_OVERCOMMIT;
+               thread_priority = reuse & WQ_FLAG_THREAD_PRIOMASK;
+               thread_newspi = 1;
+               workq = NULL;
+       } else {
+               thread_reuse = (reuse == 0)? 0: WQ_FLAG_THREAD_REUSE;
+               workq = item->workq;
+       }

 
 

-
-       workq = item->workq;
-       if (reuse == 0) {
+       
+       if (thread_reuse == 0) {

                /* reuse is set to 0, when a thread is newly created to run a workitem */
                _pthread_struct_init(self, attrs, stackaddr,  DEFAULT_STACK_SIZE, 1, 1);
                self->wqthread = 1;
                /* reuse is set to 0, when a thread is newly created to run a workitem */
                _pthread_struct_init(self, attrs, stackaddr,  DEFAULT_STACK_SIZE, 1, 1);
                self->wqthread = 1;


@@ -2917,8 +3090,13 @@ _pthread_wqthread(pthread_t self, mach_port_t kport, void * stackaddr, pthread_w
                __kdebug_trace(0x9000050, self, item, item->func_arg, 0, 0);
 #endif
                self->kernel_thread = kport;
                __kdebug_trace(0x9000050, self, item, item->func_arg, 0, 0);
 #endif
                self->kernel_thread = kport;

-               self->fun = (void *(*)(void *))item->func;
-               self->arg = item->func_arg;
+               if (thread_newspi != 0) {
+                       self->fun = (void *(*)(void *))__libdispatch_workerfunction;
+                       self->arg = thread_priority;
+               } else {
+                       self->fun = (void *(*)(void *))item->func;
+                       self->arg = item->func_arg;
+               }

                /* Add to the pthread list */
                LOCK(_pthread_list_lock);
                TAILQ_INSERT_TAIL(&__pthread_head, self, plist);
                /* Add to the pthread list */
                LOCK(_pthread_list_lock);
                TAILQ_INSERT_TAIL(&__pthread_head, self, plist);


@@ -2928,7 +3106,7 @@ _pthread_wqthread(pthread_t self, mach_port_t kport, void * stackaddr, pthread_w
                _pthread_count++;
                UNLOCK(_pthread_list_lock);
 
                _pthread_count++;
                UNLOCK(_pthread_list_lock);
 

-#if defined(__i386__) || defined(__x86_64__) || defined(__arm__)
+#if defined(__i386__) || defined(__x86_64__) || defined(__arm__) 

                if( (self->thread_id = __thread_selfid()) == (__uint64_t)-1)
                        printf("Failed to set thread_id in pthread_wqthread\n");
 #endif
                if( (self->thread_id = __thread_selfid()) == (__uint64_t)-1)
                        printf("Failed to set thread_id in pthread_wqthread\n");
 #endif


@@ -2942,8 +3120,13 @@ _pthread_wqthread(pthread_t self, mach_port_t kport, void * stackaddr, pthread_w
                if (self == NULL)
                        LIBC_ABORT("_pthread_wqthread: pthread %p setup to be NULL", self);
 
                if (self == NULL)
                        LIBC_ABORT("_pthread_wqthread: pthread %p setup to be NULL", self);
 

-               self->fun = (void *(*)(void *))item->func;
-               self->arg = item->func_arg;
+               if (thread_newspi != 0) {
+                       self->fun = (void *(*)(void *))__libdispatch_workerfunction;
+                       self->arg = NULL;
+               } else {
+                       self->fun = (void *(*)(void *))item->func;
+                       self->arg = item->func_arg;
+               }

        }
 
 #if WQ_DEBUG
        }
 
 #if WQ_DEBUG


@@ -2969,27 +3152,29 @@ _pthread_wqthread(pthread_t self, mach_port_t kport, void * stackaddr, pthread_w
        }
 #endif /* WQ_DEBUG */
 
        }
 #endif /* WQ_DEBUG */
 

-       self->cur_workq = workq;
-       self->cur_workitem = item;
-       OSAtomicDecrement32(&kernel_workq_count);
-
-       ret = (int)(intptr_t)(*self->fun)(self->arg);
+       if (thread_newspi != 0) {
+               (*__libdispatch_workerfunction)(thread_priority, thread_options, NULL);
+               _pthread_workq_return(self);
+        } else {
+               self->cur_workq = workq;
+               self->cur_workitem = item;
+               OSAtomicDecrement32Barrier(&kernel_workq_count);
+
+               ret = (int)(intptr_t)(*self->fun)(self->arg);
+               /* If we reach here without going through the above initialization path then don't go through
+               * with the teardown code path ( e.g. setjmp/longjmp ). Instead just exit this thread.
+               */
+               if (self != pthread_self()) {
+                       pthread_exit(PTHREAD_CANCELED);
+               }

 
 

-       /* If we reach here without going through the above initialization path then don't go through
-        * with the teardown code path ( e.g. setjmp/longjmp ). Instead just exit this thread.
-        */
-       if(self != pthread_self()) {
-               pthread_exit(PTHREAD_CANCELED);
+               workqueue_exit(self, workq, item);

        }
        }

-       
-       workqueue_exit(self, workq, item);
-

 }
 
 static void
 workqueue_exit(pthread_t self, pthread_workqueue_t workq, pthread_workitem_t item)
 {
 }
 
 static void
 workqueue_exit(pthread_t self, pthread_workqueue_t workq, pthread_workitem_t item)
 {

-       pthread_attr_t *attrs = &_pthread_attr_default;

        pthread_workitem_t baritem;
        pthread_workqueue_head_t headp;
        void (*func)(pthread_workqueue_t, void *);
        pthread_workitem_t baritem;
        pthread_workqueue_head_t headp;
        void (*func)(pthread_workqueue_t, void *);


@@ -3001,7 +3186,6 @@ workqueue_exit(pthread_t self, pthread_workqueue_t workq, pthread_workitem_t ite
 #if WQ_TRACE
        __kdebug_trace(0x9000070, self, 1, item->func_arg, workq->kq_count, 0);
 #endif
 #if WQ_TRACE
        __kdebug_trace(0x9000070, self, 1, item->func_arg, workq->kq_count, 0);
 #endif

-       item->flags = 0;

        free_workitem(item);
 
        if ((workq->flags & PTHREAD_WORKQ_BARRIER_ON) == PTHREAD_WORKQ_BARRIER_ON) {
        free_workitem(item);
 
        if ((workq->flags & PTHREAD_WORKQ_BARRIER_ON) == PTHREAD_WORKQ_BARRIER_ON) {


@@ -3027,7 +3211,6 @@ workqueue_exit(pthread_t self, pthread_workqueue_t workq, pthread_workitem_t ite
 #if WQ_LISTTRACE
        __kdebug_trace(0x90080a8, workq, &workq->item_listhead, workq->item_listhead.tqh_first,  workq->item_listhead.tqh_last, 0);
 #endif
 #if WQ_LISTTRACE
        __kdebug_trace(0x90080a8, workq, &workq->item_listhead, workq->item_listhead.tqh_first,  workq->item_listhead.tqh_last, 0);
 #endif

-                       baritem->flags = 0;

                        free_workitem(baritem);
                        workq->flags &= ~PTHREAD_WORKQ_BARRIER_ON;
 #if WQ_TRACE
                        free_workitem(baritem);
                        workq->flags &= ~PTHREAD_WORKQ_BARRIER_ON;
 #if WQ_TRACE


@@ -3085,6 +3268,61 @@ _pthread_workq_return(pthread_t self)
 
 /* XXXXXXXXXXXXX Pthread Workqueue functions XXXXXXXXXXXXXXXXXX */
 
 
 /* XXXXXXXXXXXXX Pthread Workqueue functions XXXXXXXXXXXXXXXXXX */
 

+int
+pthread_workqueue_setdispatch_np(void (*worker_func)(int, int, void *))
+{
+       int error = 0;
+
+       if (__workqueue_oldspis != 0)
+               return(EPERM);
+
+       __workqueue_newspis = 1;        
+
+       if (__libdispatch_workerfunction == NULL) {
+               __libdispatch_workerfunction = worker_func;
+               /* check whether the kernel supports new SPIs */
+               error = __workq_kernreturn(WQOPS_QUEUE_NEWSPISUPP, NULL, 0, 0);
+               if (error == -1){
+                       __libdispatch_workerfunction = NULL;
+                       error = ENOTSUP;
+                       __workqueue_newspis = 0;
+               } else  {
+                       /* prepare the kernel for workq action */
+                       (void)__workq_open();
+                       kernel_workq_setup = 1;
+                       if (__is_threaded == 0)
+                               __is_threaded = 1;
+                       __workqueue_newspis = 1;        
+               }
+       } else {
+               error = EBUSY;
+       }
+
+       return(error);
+}
+
+int
+pthread_workqueue_addthreads_np(int queue_priority, int options, int numthreads)
+{
+       int priority = queue_priority & WQ_FLAG_THREAD_PRIOMASK;
+       int error = 0;
+
+       /* new spi not inited yet?? */
+       if (__workqueue_newspis == 0)
+               return(EPERM);
+
+
+       if ((options & WORKQ_ADDTHREADS_OPTION_OVERCOMMIT) != 0)
+               priority |= WORKQUEUE_OVERCOMMIT;
+               
+               error = __workq_kernreturn(WQOPS_QUEUE_REQTHREADS, NULL, numthreads, priority);
+
+               if (error == -1)
+                       return(errno);
+               else
+                       return(0);
+}
+

 int 
 pthread_workqueue_create_np(pthread_workqueue_t * workqp, const pthread_workqueue_attr_t * attr)
 {
 int 
 pthread_workqueue_create_np(pthread_workqueue_t * workqp, const pthread_workqueue_attr_t * attr)
 {


@@ -3096,6 +3334,12 @@ pthread_workqueue_create_np(pthread_workqueue_t * workqp, const pthread_workqueu
                return(ENOTSUP);
        }
 #endif
                return(ENOTSUP);
        }
 #endif

+       if (__workqueue_newspis != 0)
+               return(EPERM);
+
+       if (__workqueue_oldspis == 0)
+               __workqueue_oldspis = 1;        
+

        if ((attr != NULL) && (attr->sig != PTHREAD_WORKQUEUE_ATTR_SIG)) {
                return(EINVAL);
        }
        if ((attr != NULL) && (attr->sig != PTHREAD_WORKQUEUE_ATTR_SIG)) {
                return(EINVAL);
        }


@@ -3134,6 +3378,9 @@ pthread_workqueue_additem_np(pthread_workqueue_t workq, void ( *workitem_func)(v
 {
        pthread_workitem_t witem;
 
 {
        pthread_workitem_t witem;
 

+       if (__workqueue_newspis != 0)
+               return(EPERM);
+

        if (valid_workq(workq) == 0) {
                return(EINVAL);
        }
        if (valid_workq(workq) == 0) {
                return(EINVAL);
        }


@@ -3147,10 +3394,7 @@ pthread_workqueue_additem_np(pthread_workqueue_t workq, void ( *workitem_func)(v
        witem = alloc_workitem();
        witem->func = workitem_func;
        witem->func_arg = workitem_arg;
        witem = alloc_workitem();
        witem->func = workitem_func;
        witem->func_arg = workitem_arg;

-       witem->flags = 0;

        witem->workq = workq;
        witem->workq = workq;

-       witem->item_entry.tqe_next = 0;
-       witem->item_entry.tqe_prev = 0;

 
        /* alloc workitem can drop the lock, check the state  */
        if ((workq->flags & (PTHREAD_WORKQ_IN_TERMINATE | PTHREAD_WORKQ_DESTROYED)) != 0) {
 
        /* alloc workitem can drop the lock, check the state  */
        if ((workq->flags & (PTHREAD_WORKQ_IN_TERMINATE | PTHREAD_WORKQ_DESTROYED)) != 0) {


@@ -3163,7 +3407,7 @@ pthread_workqueue_additem_np(pthread_workqueue_t workq, void ( *workitem_func)(v
        if (itemhandlep != NULL)
                *itemhandlep = (pthread_workitem_handle_t *)witem;
        if (gencountp != NULL)
        if (itemhandlep != NULL)
                *itemhandlep = (pthread_workitem_handle_t *)witem;
        if (gencountp != NULL)

-               *gencountp = witem->gencount;
+               *gencountp = 0;

 #if WQ_TRACE
        __kdebug_trace(0x9008090, witem, witem->func, witem->func_arg,  workq, 0);
 #endif
 #if WQ_TRACE
        __kdebug_trace(0x9008090, witem, witem->func, witem->func_arg,  workq, 0);
 #endif


@@ -3183,7 +3427,8 @@ pthread_workqueue_additem_np(pthread_workqueue_t workq, void ( *workitem_func)(v
 int
 pthread_workqueue_getovercommit_np(pthread_workqueue_t workq,  unsigned int *ocommp)
 {
 int
 pthread_workqueue_getovercommit_np(pthread_workqueue_t workq,  unsigned int *ocommp)
 {

-        pthread_workitem_t witem;
+       if (__workqueue_newspis != 0)
+               return(EPERM);

 
         if (valid_workq(workq) == 0) {
                 return(EINVAL);
 
         if (valid_workq(workq) == 0) {
                 return(EINVAL);


@@ -3195,16 +3440,12 @@ pthread_workqueue_getovercommit_np(pthread_workqueue_t workq,  unsigned int *oco
 }
 
 
 }
 
 

-/* DEPRECATED 
-int pthread_workqueue_removeitem_np(pthread_workqueue_t workq, pthread_workitem_handle_t itemhandle, unsigned int gencount)
-int pthread_workqueue_addbarrier_np(pthread_workqueue_t workq, void (* callback_func)(pthread_workqueue_t, void *), void * callback_arg, pthread_workitem_handle_t *itemhandlep, unsigned int *gencountp)
-int pthread_workqueue_suspend_np(pthread_workqueue_t workq)
-int pthread_workqueue_resume_np(pthread_workqueue_t workq)
-*/
-

 #else /* !BUILDING_VARIANT ] [ */
 extern int __unix_conforming;
 extern int _pthread_count;
 #else /* !BUILDING_VARIANT ] [ */
 extern int __unix_conforming;
 extern int _pthread_count;

+extern int __workqueue_newspis;
+extern int __workqueue_oldspis;
+

 extern pthread_lock_t _pthread_list_lock;
 extern void _pthread_testcancel(pthread_t thread, int isconforming);
 extern int _pthread_reap_thread(pthread_t th, mach_port_t kernel_thread, void **value_ptr, int conforming);
 extern pthread_lock_t _pthread_list_lock;
 extern void _pthread_testcancel(pthread_t thread, int isconforming);
 extern int _pthread_reap_thread(pthread_t th, mach_port_t kernel_thread, void **value_ptr, int conforming);


@@ -3220,8 +3461,7 @@ __posix_join_cleanup(void *arg)
        int already_exited, res;
        void * dummy;
        semaphore_t death;
        int already_exited, res;
        void * dummy;
        semaphore_t death;

-       mach_port_t joinport;
-       int newstyle = 0;
+       int newstyle;

 
        LOCK(thread->lock);
        already_exited = (thread->detached & _PTHREAD_EXITED);
 
        LOCK(thread->lock);
        already_exited = (thread->detached & _PTHREAD_EXITED);