Next round of fixes to the ia64 code. This includes simulated clock and

disk drivers along with a load of fixes to context switching, fork
handling and a load of other stuff I can't remember now. This takes us as
far as start_init() before it dies. I guess now I will have to finish off
the VM system and syscall handling :-).

10 files changed, 213 insertions, 147 deletions

diff --git a/sys/ia64/include/asm.h b/sys/ia64/include/asm.h
index 3e7f2f9..a0742f3 100644
--- a/sys/ia64/include/asm.h
+++ b/sys/ia64/include/asm.h
@@ -67,90 +67,48 @@
 #endif
 
 /*
- * LEAF
+ * ENTRY
  *	Declare a global leaf function.
  *	A leaf function does not call other functions.
  */
-#define	LEAF(_name_, _n_args_)			\
+#define	ENTRY(_name_, _n_args_)			\
 	.global	_name_;				\
+	.align	16;				\
 	.proc	_name_;				\
 _name_:;					\
 	.regstk	_n_args_, 0, 0, 0		\
 	MCOUNT
 
-#define	LEAF_NOPROFILE(_name_, _n_args_)	\
+#define	ENTRY_NOPROFILE(_name_, _n_args_)	\
 	.global	_name_;				\
+	.align	16;				\
 	.proc	_name_;				\
 _name_:;					\
 	.regstk	_n_args_, 0, 0, 0
 
 /*
- * STATIC_LEAF
+ * STATIC_ENTRY
  *	Declare a local leaf function.
  */
-#define STATIC_LEAF(_name_, _n_args_)		\
+#define STATIC_ENTRY(_name_, _n_args_)		\
+	.align	16;				\
 	.proc	_name_;				\
 _name_:;					\
 	.regstk	_n_args_, 0, 0, 0		\
 	MCOUNT
 /*
- * XLEAF
+ * XENTRY
  *	Global alias for a leaf function, or alternate entry point
  */
-#define	XLEAF(_name_)				\
+#define	XENTRY(_name_)				\
 	.globl	_name_;				\
 _name_:
 
 /*
- * STATIC_XLEAF
+ * STATIC_XENTRY
  *	Local alias for a leaf function, or alternate entry point
  */
-#define	STATIC_XLEAF(_name_)			\
-_name_:
-
-/*
- * NESTED
- *	Declare a (global) nested function
- *	A nested function calls other functions and needs
- *	to use alloc to save registers.
- */
-#define	NESTED(_name_,_n_args_)			\
-	.globl	_name_;				\
-	.proc	_name_;				\
-_name_:;					\
-	.regstk	_n_args_, 0, 0, 0		\
-	MCOUNT
-
-#define	NESTED_NOPROFILE(_name_,_n_args_)	\
-	.globl	_name_;				\
-	.proc	_name_;				\
-_name_:;					\
-	.regstk	_n_args_, 0, 0, 0
-
-/*
- * STATIC_NESTED
- *	Declare a local nested function.
- */
-#define	STATIC_NESTED(_name_,_n_args_)		\
-	.proc	_name_;				\
-_name_:;					\
-	.regstk	_n_args_, 0, 0, 0		\
-	MCOUNT
-
-/*
- * XNESTED
- *	Same as XLEAF, for a nested function.
- */
-#define	XNESTED(_name_)				\
-	.globl	_name_;				\
-_name_:
-
-
-/*
- * STATIC_XNESTED
- *	Same as STATIC_XLEAF, for a nested function.
- */
-#define	STATIC_XNESTED(_name_)			\
+#define	STATIC_XENTRY(_name_)			\
 _name_:
 
 
@@ -201,10 +159,11 @@ _name_	=	_value_
  *	Allocate space for a message (a read-only ascii string)
  */
 #define	ASCIZ	.asciz
-#define	MSG(msg,reg,label)					\
-	lda reg, label;						\
-	.data;							\
-label:	ASCIZ msg;						\
+#define	MSG(msg,reg,label)			\
+	addl reg,@ltoff(label),gp;;		\
+	ld8 reg=[reg];;				\
+	.data;					\
+label:	ASCIZ msg;				\
 	.text;
 
 /*
diff --git a/sys/ia64/include/ia64_cpu.h b/sys/ia64/include/ia64_cpu.h
index 4de9938..1639e88 100644
--- a/sys/ia64/include/ia64_cpu.h
+++ b/sys/ia64/include/ia64_cpu.h
@@ -183,6 +183,166 @@ ia64_tpa(u_int64_t va)
 }
 
 /*
+ * Read the value of ar.k0.
+ */
+static __inline u_int64_t
+ia64_get_k0(void)
+{
+	u_int64_t result;
+	__asm __volatile("mov %0=ar.k0" : "=r" (result));
+	return result;
+}
+
+/*
+ * Read the value of ar.k1.
+ */
+static __inline u_int64_t
+ia64_get_k1(void)
+{
+	u_int64_t result;
+	__asm __volatile("mov %0=ar.k1" : "=r" (result));
+	return result;
+}
+
+/*
+ * Read the value of ar.k2.
+ */
+static __inline u_int64_t
+ia64_get_k2(void)
+{
+	u_int64_t result;
+	__asm __volatile("mov %0=ar.k2" : "=r" (result));
+	return result;
+}
+
+/*
+ * Read the value of ar.k3.
+ */
+static __inline u_int64_t
+ia64_get_k3(void)
+{
+	u_int64_t result;
+	__asm __volatile("mov %0=ar.k3" : "=r" (result));
+	return result;
+}
+
+/*
+ * Read the value of ar.k4.
+ */
+static __inline u_int64_t
+ia64_get_k4(void)
+{
+	u_int64_t result;
+	__asm __volatile("mov %0=ar.k4" : "=r" (result));
+	return result;
+}
+
+/*
+ * Read the value of ar.k5.
+ */
+static __inline u_int64_t
+ia64_get_k5(void)
+{
+	u_int64_t result;
+	__asm __volatile("mov %0=ar.k5" : "=r" (result));
+	return result;
+}
+
+/*
+ * Read the value of ar.k6.
+ */
+static __inline u_int64_t
+ia64_get_k6(void)
+{
+	u_int64_t result;
+	__asm __volatile("mov %0=ar.k6" : "=r" (result));
+	return result;
+}
+
+/*
+ * Read the value of ar.k7.
+ */
+static __inline u_int64_t
+ia64_get_k7(void)
+{
+	u_int64_t result;
+	__asm __volatile("mov %0=ar.k7" : "=r" (result));
+	return result;
+}
+
+/*
+ * Write the value of ar.k0.
+ */
+static __inline void
+ia64_set_k0(u_int64_t v)
+{
+	__asm __volatile("mov ar.k0=%0" :: "r" (v));
+}
+
+/*
+ * Write the value of ar.k1.
+ */
+static __inline void
+ia64_set_k1(u_int64_t v)
+{
+	__asm __volatile("mov ar.k1=%0" :: "r" (v));
+}
+
+/*
+ * Write the value of ar.k2.
+ */
+static __inline void
+ia64_set_k2(u_int64_t v)
+{
+	__asm __volatile("mov ar.k2=%0" :: "r" (v));
+}
+
+/*
+ * Write the value of ar.k3.
+ */
+static __inline void
+ia64_set_k3(u_int64_t v)
+{
+	__asm __volatile("mov ar.k3=%0" :: "r" (v));
+}
+
+/*
+ * Write the value of ar.k4.
+ */
+static __inline void
+ia64_set_k4(u_int64_t v)
+{
+	__asm __volatile("mov ar.k4=%0" :: "r" (v));
+}
+
+/*
+ * Write the value of ar.k5.
+ */
+static __inline void
+ia64_set_k5(u_int64_t v)
+{
+	__asm __volatile("mov ar.k5=%0" :: "r" (v));
+}
+
+/*
+ * Write the value of ar.k6.
+ */
+static __inline void
+ia64_set_k6(u_int64_t v)
+{
+	__asm __volatile("mov ar.k6=%0" :: "r" (v));
+}
+
+/*
+ * Write the value of ar.k7.
+ */
+static __inline void
+ia64_set_k7(u_int64_t v)
+{
+	__asm __volatile("mov ar.k7=%0" :: "r" (v));
+}
+
+/*
  * Read the value of ar.itc.
  */
 static __inline u_int64_t
diff --git a/sys/ia64/include/mutex.h b/sys/ia64/include/mutex.h
index b858079..ea81a8b 100644
--- a/sys/ia64/include/mutex.h
+++ b/sys/ia64/include/mutex.h
@@ -536,7 +536,9 @@ _mtx_exit(struct mtx *mtxp, int type, const char *file, int line)
 #define MTX_ENTER(lck, rPSR, rOLD, rNEW, rLCK)		\
 	mov	rPSR=psr ;				\
 	mov	rNEW=globalp ;				\
-	mov	rLCK=lck+MTX_LOCK ;;			\
+	addl	rLCK=@ltoff(lck),gp ;;			\
+	ld8	rLCK=[rLCK] ;;				\
+	add	rLCK=MTX_LOCK,rLCK ;;			\
 	rsm	psr.i ;					\
 	mov	ar.ccv=MTX_UNOWNED ;			\
 	add	rNEW=PC_CURPROC,rNEW ;;			\
@@ -544,18 +546,19 @@ _mtx_exit(struct mtx *mtxp, int type, const char *file, int line)
 1:	cmpxchg8.acq rOLD=[rLCK],rNEW,ar.ccv ;;		\
 	cmp.eq	p1,p0=MTX_UNOWNED,rOLD ;;		\
 (p1)	br.cond.spnt.few 1b ;;				\
-	mov	rLCK=lck+MTX_SAVEPSR ;;			\
+	addl	rLCK=@ltoff(lck),gp ;;			\
+	ld8	rLCK=[rLCK] ;;				\
+	add	rLCK=MTX_SAVEPSR,rLCK ;;		\
 	st4	[rLCK]=rPSR
 
-#define MTX_EXIT(lck, rTMP, rLCK)		\
-	mov	rTMP=MTX_UNOWNED ;		\
-	addl	rLCK=@ltoff(lck),gp;;		\
-	add	rLCK=MTX_LOCK,rLCK;;		\
-	st8.rel	[rLCK]=rTMP ;;			\
-	addl	rLCK=@ltoff(lck),gp;;		\
-	add	rLCK=MTX_SAVEPSR,rLCK;;		\
-	ld4	rTMP=[rLCK] ;;			\
-	mov	psr.l=rTMP ;;			\
+#define MTX_EXIT(lck, rTMP, rLCK)			\
+	mov	rTMP=MTX_UNOWNED ;			\
+	addl	rLCK=@ltoff(lck),gp;;			\
+	ld8	rLCK=[rLCK];;				\
+	add	rLCK=MTX_LOCK,rLCK;;			\
+	st8.rel	[rLCK]=rTMP,MTX_SAVEPSR-MTX_LOCK ;;	\
+	ld4	rTMP=[rLCK] ;;				\
+	mov	psr.l=rTMP ;;				\
 	srlz.d
 
 #endif	/* !LOCORE */
diff --git a/sys/ia64/include/pal.h b/sys/ia64/include/pal.h
index 39c12a4..f492608 100644
--- a/sys/ia64/include/pal.h
+++ b/sys/ia64/include/pal.h
@@ -88,8 +88,8 @@ struct ia64_pal_result {
 };
 
 extern struct ia64_pal_result
-		ia64_call_pal_static(u_int64_t proc, u_int64_t arg1,
-				     u_int64_t arg2, u_int64_t arg3);
+	ia64_call_pal(u_int64_t proc, u_int64_t arg1,
+		      u_int64_t arg2, u_int64_t arg3);
 
 
 #endif /* _MACHINE_PAL_H_ */
diff --git a/sys/ia64/include/pcb.h b/sys/ia64/include/pcb.h
index d8679ca..49a5c54 100644
--- a/sys/ia64/include/pcb.h
+++ b/sys/ia64/include/pcb.h
@@ -58,6 +58,7 @@ struct pcb {
 	u_int64_t		pcb_unat;	/* ar.unat for r4..r7 */
 	u_int64_t		pcb_rnat;
 	u_int64_t		pcb_pr;		/* predicates */
+	u_int64_t		pcb_schednest;	/* sched_lock state */
 
 	u_int64_t		pcb_onfault;	/* for copy faults */
 	u_int64_t		pcb_accessaddr;	/* for [fs]uswintr */
diff --git a/sys/ia64/include/proc.h b/sys/ia64/include/proc.h
index 226aa94..4aa8566 100644
--- a/sys/ia64/include/proc.h
+++ b/sys/ia64/include/proc.h
@@ -37,6 +37,7 @@
 
 struct mdproc {
 	u_long		md_flags;
+	vm_offset_t	md_bspstore;	/* initial ar.bspstore */
 	struct	trapframe *md_tf;	/* trap/syscall registers */
 };
 
diff --git a/sys/ia64/include/profile.h b/sys/ia64/include/profile.h
index 57fbb69..70a9cf7 100644
--- a/sys/ia64/include/profile.h
+++ b/sys/ia64/include/profile.h
@@ -136,74 +136,8 @@ LX99:	SETGP(pv)
 
 #define MCOUNT __asm ("		\
 	.globl	_mcount;	\
-	.ent	_mcount 0;	\
+	.proc	_mcount;	\
 _mcount:;			\
-	.frame	$30,0,$26;	\
-	.set noat;		\
-	.set noreorder;		\
-				\
-	lda	$30, -192($30);	\
-				\
-	stq	$28, 0($30);	\
-	stq	$0, 8($30);	\
-	stq	$1, 16($30);	\
-	stq	$2, 24($30);	\
-	stq	$3, 32($30);	\
-	stq	$4, 40($30);	\
-	stq	$5, 48($30);	\
-	stq	$6, 56($30);	\
-	stq	$7, 64($30);	\
-	stq	$8, 72($30);	\
-	stq	$15, 80($30);	\
-	stq	$16, 88($30);	\
-	stq	$17, 96($30);	\
-	stq	$18, 104($30);	\
-	stq	$19, 112($30);	\
-	stq	$20, 120($30);	\
-	stq	$21, 128($30);	\
-	stq	$22, 136($30);	\
-	stq	$23, 144($30);	\
-	stq	$24, 152($30);	\
-	stq	$25, 160($30);	\
-	stq	$26, 168($30);	\
-	stq	$27, 176($30);	\
-	stq	$29, 184($30);	\
-				\
-	br	$27, LX98;	\
-LX98:	ldgp	$29,0($27);	\
-	mov	$26, $16;	\
-	mov	$28, $17;	\
-	jsr	$26,mcount;	\
-	ldgp	$29,0($26);	\
-				\
-	ldq	$0, 8($30);	\
-	ldq	$1, 16($30);	\
-	ldq	$2, 24($30);	\
-	ldq	$3, 32($30);	\
-	ldq	$4, 40($30);	\
-	ldq	$5, 48($30);	\
-	ldq	$6, 56($30);	\
-	ldq	$7, 64($30);	\
-	ldq	$8, 72($30);	\
-	ldq	$15, 80($30);	\
-	ldq	$16, 88($30);	\
-	ldq	$17, 96($30);	\
-	ldq	$18, 104($30);	\
-	ldq	$19, 112($30);	\
-	ldq	$20, 120($30);	\
-	ldq	$21, 128($30);	\
-	ldq	$22, 136($30);	\
-	ldq	$23, 144($30);	\
-	ldq	$24, 152($30);	\
-	ldq	$25, 160($30);	\
-	ldq	$26, 168($30);	\
-	ldq	$27, 176($30);	\
-	ldq	$29, 184($30);	\
-				\
-	ldq	$28, 0($30);	\
-				\
-	lda	$30, 192($30);	\
-	ret	$31, ($28), 1;	\
 				\
 	.end	_mcount");
 
diff --git a/sys/ia64/include/reg.h b/sys/ia64/include/reg.h
index 7378e37..302cd6a 100644
--- a/sys/ia64/include/reg.h
+++ b/sys/ia64/include/reg.h
@@ -29,10 +29,16 @@
 #ifndef _MACHINE_REG_H_
 #define _MACHINE_REG_H_
 
+#ifndef _IA64_FPREG_DEFINED
+
 struct ia64_fpreg {
 	u_int64_t	fpr_bits[2];
 } __attribute__ ((aligned (16)));
 
+#define _IA64_FPREG_DEFINED
+
+#endif
+
 struct reg {
 	u_int64_t	r_regs[128];
 };
diff --git a/sys/ia64/include/signal.h b/sys/ia64/include/signal.h
index 5474941..006a99b 100644
--- a/sys/ia64/include/signal.h
+++ b/sys/ia64/include/signal.h
@@ -35,6 +35,16 @@ typedef long	sig_atomic_t;
 
 #ifndef _ANSI_SOURCE
 
+#ifndef _IA64_FPREG_DEFINED
+
+struct ia64_fpreg {
+	unsigned long	fpr_bits[2];
+} __attribute__ ((aligned (16)));
+
+#define _IA64_FPREG_DEFINED
+
+#endif
+
 /*
  * Information pushed on stack when a signal is delivered.
  * This is used by the kernel to restore state following
diff --git a/sys/ia64/include/vmparam.h b/sys/ia64/include/vmparam.h
index 2cdc959..1c9c245 100644
--- a/sys/ia64/include/vmparam.h
+++ b/sys/ia64/include/vmparam.h
@@ -54,15 +54,7 @@
  * kernel stack.
  */
 #define	USRTEXT		CLBYTES
-/* #define	USRSTACK	VM_MAXUSER_ADDRESS */
-
-/*
- * This stack location is suitable for OSF1 emulation.  Some OSF
- * programs are built as 32bit and assume that the stack is reachable
- * with a 32bit value.  OSF1 manages to have a variable location for
- * the user stack which we should probably also support.
- */
-#define USRSTACK	(0x12000000LL - (UPAGES*PAGE_SIZE))
+#define	USRSTACK	VM_MAXUSER_ADDRESS
 
 /*
  * Virtual memory related constants, all in bytes