Move the mailbox to the beginning of the thread and align the

thread so that the context (SSE FPU state) is also aligned.

10 files changed, 48 insertions, 14 deletions

diff --git a/lib/libkse/arch/i386/include/pthread_md.h b/lib/libkse/arch/i386/include/pthread_md.h
index 2611101..cb5a344 100644
--- a/lib/libkse/arch/i386/include/pthread_md.h
+++ b/lib/libkse/arch/i386/include/pthread_md.h
@@ -48,4 +48,7 @@ extern int _thr_getcontext(ucontext_t *);
 
 #define	THR_GETCONTEXT(ucp)	_thr_getcontext(ucp)
 #define	THR_SETCONTEXT(ucp)	_thr_setcontext(ucp)
+
+#define	THR_ALIGNBYTES	15
+#define	THR_ALIGN(td)	(((unsigned)(td) + THR_ALIGNBYTES) & ~THR_ALIGNBYTES)
 #endif
diff --git a/lib/libkse/thread/thr_create.c b/lib/libkse/thread/thr_create.c
index 4c65d3c..0c9edbd 100644
--- a/lib/libkse/thread/thr_create.c
+++ b/lib/libkse/thread/thr_create.c
@@ -97,6 +97,7 @@ _pthread_create(pthread_t * thread, const pthread_attr_t * attr,
 	struct pthread *curthread, *new_thread;
 	struct kse *kse = NULL;
 	struct kse_group *kseg = NULL;
+	void *p;
 	kse_critical_t crit;
 	int i;
 	int ret = 0;
@@ -123,7 +124,9 @@ _pthread_create(pthread_t * thread, const pthread_attr_t * attr,
 		ret = EAGAIN;
 	} else {
 		/* Initialize the thread structure: */
+		p = new_thread->alloc_addr;
 		memset(new_thread, 0, sizeof(struct pthread));
+		new_thread->alloc_addr = p;
 
 		/* Check if default thread attributes are required: */
 		if (attr == NULL || *attr == NULL)
diff --git a/lib/libkse/thread/thr_init.c b/lib/libkse/thread/thr_init.c
index d43adf5..f06df6c 100644
--- a/lib/libkse/thread/thr_init.c
+++ b/lib/libkse/thread/thr_init.c
@@ -311,10 +311,13 @@ _libpthread_init(struct pthread *curthread)
 static void
 init_main_thread(struct pthread *thread)
 {
+	void *p;
 	int i;
 
 	/* Zero the initial thread structure. */
+	p = thread->alloc_addr;
 	memset(thread, 0, sizeof(struct pthread));
+	thread->alloc_addr = p;
 
 	/* Setup the thread attributes. */
 	thread->attr = _pthread_attr_default;
diff --git a/lib/libkse/thread/thr_kern.c b/lib/libkse/thread/thr_kern.c
index a153a4d..54d9dd7 100644
--- a/lib/libkse/thread/thr_kern.c
+++ b/lib/libkse/thread/thr_kern.c
@@ -2019,6 +2019,7 @@ struct pthread *
 _thr_alloc(struct pthread *curthread)
 {
 	kse_critical_t crit;
+	void *p;
 	struct pthread *thread = NULL;
 
 	if (curthread != NULL) {
@@ -2035,8 +2036,13 @@ _thr_alloc(struct pthread *curthread)
 			_kse_critical_leave(crit);
 		}
 	}
-	if (thread == NULL)
-		thread = (struct pthread *)malloc(sizeof(struct pthread));
+	if (thread == NULL) {
+		p = malloc(sizeof(struct pthread) + THR_ALIGNBYTES);
+		if (p != NULL) {
+			thread = (struct pthread *)THR_ALIGN(p);
+			thread->alloc_addr = p;
+		}
+	}
 	return (thread);
 }
 
@@ -2052,7 +2058,7 @@ _thr_free(struct pthread *curthread, struct pthread *thread)
 			_lockuser_destroy(&thread->lockusers[i]);
 		}
 		_lock_destroy(&thread->lock);
-		free(thread);
+		free(thread->alloc_addr);
 	}
 	else {
 		crit = _kse_critical_enter();
diff --git a/lib/libkse/thread/thr_private.h b/lib/libkse/thread/thr_private.h
index 4d43242..22d2445 100644
--- a/lib/libkse/thread/thr_private.h
+++ b/lib/libkse/thread/thr_private.h
@@ -593,6 +593,12 @@ struct pthread_specific_elem {
  */
 struct pthread {
 	/*
+	 * Thread mailbox is first so it cal be aligned properly.
+	 */
+	struct kse_thr_mailbox	tmbx;
+	void			*alloc_addr;	/* real address (unaligned) */
+
+	/*
 	 * Magic value to help recognize a valid thread structure
 	 * from an invalid one:
 	 */
@@ -626,10 +632,6 @@ struct pthread {
 	void			*arg;
 	struct pthread_attr	attr;
 
-	/*
-	 * Thread mailbox.
-	 */
-	struct kse_thr_mailbox	tmbx;
 	int			active;		/* thread running */
 	int			blocked;	/* thread blocked in kernel */
 	int			need_switchout;
diff --git a/lib/libpthread/arch/i386/include/pthread_md.h b/lib/libpthread/arch/i386/include/pthread_md.h
index 2611101..cb5a344 100644
--- a/lib/libpthread/arch/i386/include/pthread_md.h
+++ b/lib/libpthread/arch/i386/include/pthread_md.h
@@ -48,4 +48,7 @@ extern int _thr_getcontext(ucontext_t *);
 
 #define	THR_GETCONTEXT(ucp)	_thr_getcontext(ucp)
 #define	THR_SETCONTEXT(ucp)	_thr_setcontext(ucp)
+
+#define	THR_ALIGNBYTES	15
+#define	THR_ALIGN(td)	(((unsigned)(td) + THR_ALIGNBYTES) & ~THR_ALIGNBYTES)
 #endif
diff --git a/lib/libpthread/thread/thr_create.c b/lib/libpthread/thread/thr_create.c
index 4c65d3c..0c9edbd 100644
--- a/lib/libpthread/thread/thr_create.c
+++ b/lib/libpthread/thread/thr_create.c
@@ -97,6 +97,7 @@ _pthread_create(pthread_t * thread, const pthread_attr_t * attr,
 	struct pthread *curthread, *new_thread;
 	struct kse *kse = NULL;
 	struct kse_group *kseg = NULL;
+	void *p;
 	kse_critical_t crit;
 	int i;
 	int ret = 0;
@@ -123,7 +124,9 @@ _pthread_create(pthread_t * thread, const pthread_attr_t * attr,
 		ret = EAGAIN;
 	} else {
 		/* Initialize the thread structure: */
+		p = new_thread->alloc_addr;
 		memset(new_thread, 0, sizeof(struct pthread));
+		new_thread->alloc_addr = p;
 
 		/* Check if default thread attributes are required: */
 		if (attr == NULL || *attr == NULL)
diff --git a/lib/libpthread/thread/thr_init.c b/lib/libpthread/thread/thr_init.c
index d43adf5..f06df6c 100644
--- a/lib/libpthread/thread/thr_init.c
+++ b/lib/libpthread/thread/thr_init.c
@@ -311,10 +311,13 @@ _libpthread_init(struct pthread *curthread)
 static void
 init_main_thread(struct pthread *thread)
 {
+	void *p;
 	int i;
 
 	/* Zero the initial thread structure. */
+	p = thread->alloc_addr;
 	memset(thread, 0, sizeof(struct pthread));
+	thread->alloc_addr = p;
 
 	/* Setup the thread attributes. */
 	thread->attr = _pthread_attr_default;
diff --git a/lib/libpthread/thread/thr_kern.c b/lib/libpthread/thread/thr_kern.c
index a153a4d..54d9dd7 100644
--- a/lib/libpthread/thread/thr_kern.c
+++ b/lib/libpthread/thread/thr_kern.c
@@ -2019,6 +2019,7 @@ struct pthread *
 _thr_alloc(struct pthread *curthread)
 {
 	kse_critical_t crit;
+	void *p;
 	struct pthread *thread = NULL;
 
 	if (curthread != NULL) {
@@ -2035,8 +2036,13 @@ _thr_alloc(struct pthread *curthread)
 			_kse_critical_leave(crit);
 		}
 	}
-	if (thread == NULL)
-		thread = (struct pthread *)malloc(sizeof(struct pthread));
+	if (thread == NULL) {
+		p = malloc(sizeof(struct pthread) + THR_ALIGNBYTES);
+		if (p != NULL) {
+			thread = (struct pthread *)THR_ALIGN(p);
+			thread->alloc_addr = p;
+		}
+	}
 	return (thread);
 }
 
@@ -2052,7 +2058,7 @@ _thr_free(struct pthread *curthread, struct pthread *thread)
 			_lockuser_destroy(&thread->lockusers[i]);
 		}
 		_lock_destroy(&thread->lock);
-		free(thread);
+		free(thread->alloc_addr);
 	}
 	else {
 		crit = _kse_critical_enter();
diff --git a/lib/libpthread/thread/thr_private.h b/lib/libpthread/thread/thr_private.h
index 4d43242..22d2445 100644
--- a/lib/libpthread/thread/thr_private.h
+++ b/lib/libpthread/thread/thr_private.h
@@ -593,6 +593,12 @@ struct pthread_specific_elem {
  */
 struct pthread {
 	/*
+	 * Thread mailbox is first so it cal be aligned properly.
+	 */
+	struct kse_thr_mailbox	tmbx;
+	void			*alloc_addr;	/* real address (unaligned) */
+
+	/*
 	 * Magic value to help recognize a valid thread structure
 	 * from an invalid one:
 	 */
@@ -626,10 +632,6 @@ struct pthread {
 	void			*arg;
 	struct pthread_attr	attr;
 
-	/*
-	 * Thread mailbox.
-	 */
-	struct kse_thr_mailbox	tmbx;
 	int			active;		/* thread running */
 	int			blocked;	/* thread blocked in kernel */
 	int			need_switchout;