From 2c3633d2e20d276322394326adff8f5d4a0466ec Mon Sep 17 00:00:00 2001
From: David du Colombier <0intro@gmail.com>
Date: Fri, 29 Sep 2000 00:00:00 +0000
Subject: [PATCH] Plan 9 from Bell Labs 2000-09-29

---
 alphapc/main.c |  3 ++-
 bitsy/dat.h    |  9 ++++---
 bitsy/fns.h    |  7 ++++--
 bitsy/l.s      | 28 +++++++++++++++++----
 bitsy/main.c   |  3 +++
 bitsy/mem.h    |  3 ++-
 bitsy/mmu.c    | 66 +++++++++++++++++++++++++++++++-------------------
 bitsy/trap.c   | 23 +++++++++++++++++-
 carrera/l.s    |  2 ++
 pc/l.s         |  6 +++--
 10 files changed, 109 insertions(+), 41 deletions(-)

diff --git a/alphapc/main.c b/alphapc/main.c
index 6799c1700045339bb2c0a26d32a8300723f5b688..7295870555d155c0dadb1815e748de5388e8d4df 100644
--- a/alphapc/main.c
+++ b/alphapc/main.c
@@ -15,7 +15,8 @@ Hwrpb *hwrpb;
 Bootconf *bootconf;
 Conf	conf;
 FPsave	initfp;
-uvlong initfpcr = 0x2800800000000000LL;
+	/* setfcr(FPPDBL|FPRNR|FPINVAL|FPZDIV|FPOVFL) */
+uvlong initfpcr = (1LL<62)|(1LL<61)|(1LL<60)|(2LL<<58)|(1LL<48);
 
 char bootargs[BOOTARGSLEN];
 char *confname[MAXCONF];
diff --git a/bitsy/dat.h b/bitsy/dat.h
index 6b49b222c34292be481904b5d2e37d30cd36c43b..9699990855754788b1d7ec0dbd4c49b126786443 100644
--- a/bitsy/dat.h
+++ b/bitsy/dat.h
@@ -167,10 +167,10 @@ struct
 }active;
 
 /*
- * Each processor sees its own Mach structure at address MACHADDR.
- * However, the Mach structures must also be available via the per-processor
- * MMU information array machp, mainly for disambiguation and access to
- * the clock which is only maintained by the bootstrap processor (0).
+ *  Each processor sees its own Mach structure at address MACHADDR.
+ *  However, the Mach structures must also be available via the per-processor
+ *  MMU information array machp, mainly for disambiguation and access to
+ *  the clock which is only maintained by the bootstrap processor (0).
  */
 Mach* machp[MAXMACH];
 	
@@ -183,3 +183,4 @@ enum
 {
 	OneMeg=	1024*1024,
 };
+
diff --git a/bitsy/fns.h b/bitsy/fns.h
index cc95eb0481e6cc0566d9a2fdc578c7daa4128a91..0d1b0267ad8959f87303f69d27cab9803f1b5de2 100644
--- a/bitsy/fns.h
+++ b/bitsy/fns.h
@@ -5,17 +5,21 @@ int	cistrncmp(char*, char*, int);
 #define	clearmmucache()				/* x86 doesn't have one */
 void	clockintr(Ureg*, void*);
 void	clockintrsched(void);
-void	(*coherence)(void);
+#define	coherence()
 void	delay(int);
 void	evenaddr(ulong);
+void	exceptionvectors(void);
 ulong	getfar(void);
 ulong	getfsr(void);
 #define	getpgcolor(a)	0
 void	idle(void);
 #define	idlehands()			/* nothing to do in the runproc */
 int	iprint(char*, ...);
+ulong*	mapspecial(ulong, int);
 void	meminit(void);
 void	mmuinit(void);
+void	mmuenable(void);
+void	mmudisable(void);
 #define	procrestore(p)
 void	procsave(Proc*);
 void	procsetup(Proc*);
@@ -26,7 +30,6 @@ void	screeninit(void);
 int	screenprint(char*, ...);			/* debugging */
 void	(*screenputs)(char*, int);
 void	touser(void*);
-void	trapenable(int, void (*)(Ureg*, void*), void*, char*);
 void	trapinit(void);
 int	tas(void*);
 void	wbflush(void);
diff --git a/bitsy/l.s b/bitsy/l.s
index 9bb7f507dce63dbca664a1b277bbb38c6a7927b3..d05c03906fde423e07f2ce3ad845aa0f0755076c 100644
--- a/bitsy/l.s
+++ b/bitsy/l.s
@@ -65,13 +65,13 @@ TEXT flushcache(SB), $-4
 	RET
 
 /* drain write buffer */
-TEXT drainwb(SB), $-4
+TEXT wbflush(SB), $-4
 	MCR	CpMMU, 0, R0, C(CpCacheFlush), C(0x0), 4
 	RET
 
 /* return cpu id */
 TEXT getcpuid(SB), $-4
-	MRC	CpMMU, 0, R0, C(CpControl), C(0x0)
+	MRC	CpMMU, 0, R0, C(CpCPUID), C(0x0)
 	RET
 
 /* return fault status */
@@ -87,14 +87,32 @@ TEXT getfar(SB), $-4
 /* set the translation table base */
 TEXT putttb(SB), $-4
 	MCR	CpMMU, 0, R0, C(CpTTB), C(0x0)
+	RET
+
+/*
+ *  enable mmu, i and d caches, and exception vectors at 0xffff0000
+ */
+TEXT mmuenable(SB), $-4
+	MRC	CpMMU, 0, R0, C(CpControl), C(0x0)
+	ORR	$(CpCmmuena|CpCdcache|CpCicache|CpCvivec), R0
+	MCR     CpMMU, 0, R0, C(CpControl), C(0x0)
+	RET
+
+TEXT mmudisable(SB), $-4
+	MRC	CpMMU, 0, R0, C(CpControl), C(0x0)
+	BIC	$(CpCmmuena|CpCdcache|CpCicache|CpCvivec), R0
+	MCR     CpMMU, 0, R0, C(CpControl), C(0x0)
+	RET
 
 /* set the translation table base */
 TEXT putdac(SB), $-4
 	MCR	CpMMU, 0, R0, C(CpDAC), C(0x0)
+	RET
 
-/* set the translation table base */
+/* set address translation pid */
 TEXT putpid(SB), $-4
 	MCR	CpMMU, 0, R0, C(CpPID), C(0x0)
+	RET
 
 /*
  *  set the stack value for the mode passed in R0
@@ -116,7 +134,7 @@ TEXT setr13(SB), $-4
 /*
  *  exception vectors, copied by trapinit() to somewhere useful
  */
-TEXT vectors(SB), $-4
+TEXT exceptionvectors(SB), $-4
 	MOVW	0x18(R15), R15		/* reset */
 	MOVW	0x18(R15), R15		/* undefined */
 	MOVW	0x18(R15), R15		/* SWI */
@@ -206,7 +224,7 @@ _vexcep:
 	MOVW	$setR12(SB), R12	/* Make sure we've got the kernel's SB loaded */
 	MOVW	R13, R0			/* first arg is pointer to ureg */
 	SUB	$8, R13			/* space for argument+link */
-	BL	exception(SB)
+	BL	trap(SB)
 
 _vrfe: 
 	ADD	$(8+4*15), R13		/* make r13 point to ureg->type */
diff --git a/bitsy/main.c b/bitsy/main.c
index 2b15de499c27daabefbc529fac590e28a34867a0..59e0035db5d3d6d0bbc8a5af013021600dcf9ddb 100644
--- a/bitsy/main.c
+++ b/bitsy/main.c
@@ -24,6 +24,8 @@ main(void)
 	iprint("after xinit\n");
 	mmuinit();
 	iprint("after mmuinit\n");
+	trapinit();
+	iprint("after trapinit\n");
 	delay(100000);
 	exit(1);
 }
@@ -39,6 +41,7 @@ exit(int ispanic)
 	USED(ispanic);
 	delay(1000);
 
+	mmudisable();
 	f = nil;
 	(*f)();
 }
diff --git a/bitsy/mem.h b/bitsy/mem.h
index 3f4a835b219ab8fb2530d2c786a0dfa8247b40a2..9f6158297476dea12856c357b83efa9a05b36884 100644
--- a/bitsy/mem.h
+++ b/bitsy/mem.h
@@ -53,6 +53,7 @@
 #define	TSTKTOP		(USTKTOP-USTKSIZE)	/* end of new stack in sysexec */
 #define TSTKSIZ 	100
 #define MACHADDR	(KZERO+0x00001000)
+#define	EVECTORS	0xFFFF0000		/* base of exception vectors (also mapped to KZERO) */
 
 #define KSTACK		(16*1024)		/* Size of kernel stack */
 
@@ -134,7 +135,7 @@
 /*
  *  CpControl
  */
-#define CpCmmu		0x00000001	/* M: MMU enable */
+#define CpCmmuena	0x00000001	/* M: MMU enable */
 #define CpCalign	0x00000002	/* A: alignment fault enable */
 #define CpCdcache	0x00000004	/* C: data cache on */
 #define CpCwb		0x00000008	/* W: write buffer turned on */
diff --git a/bitsy/mmu.c b/bitsy/mmu.c
index 9d794570ed1ae54de51f4214b328e29daa1496ba..677deb4e0bcf9c9c148ec2fb6ab355ea24b296a6 100644
--- a/bitsy/mmu.c
+++ b/bitsy/mmu.c
@@ -62,6 +62,7 @@ void
 mmuinit(void)
 {
 	ulong a, e;
+	ulong *t;
 
 	/* get a prototype level 1 page */
 	l1table = xspanalloc(BY2PG, 16*1024, 0);
@@ -70,33 +71,39 @@ mmuinit(void)
 	/* direct map DRAM */
 	e = conf.base1 + BY2PG*conf.npage1;
 	for(a = PHYSDRAM0; a < e; a += OneMeg)
-		l1table[a>>20] = L1Section | L1KernelRW |
-				L1Cached | L1Buffered | (a&L1SectBaseMask);
-
-	/* direct map devs */
-	for(a = REGZERO; a < REGTOP; a += OneMeg)
-		l1table[a>>20] = L1Section | L1KernelRW | (a&L1SectBaseMask);
+		l1table[a>>20] = L1Section | L1KernelRW | (a&L1SectBaseMask) |
+				L1Cached | L1Buffered;
 
 	/* direct map zeros area */
 	for(a = PHYSNULL0; a < PHYSNULL0 + 128 * OneMeg; a += OneMeg)
-		l1table[a>>20] = L1Section | L1KernelRW |
-				L1Cached | L1Buffered | (a&L1SectBaseMask);
+		l1table[a>>20] = L1Section | L1KernelRW | (a&L1SectBaseMask);
 
 	/* direct map flash */
 	for(a = PHYSFLASH0; a < PHYSFLASH0 + 128 * OneMeg; a += OneMeg)
-		l1table[a>>20] = L1Section | L1KernelRW |
-				L1Cached | L1Buffered | (a&L1SectBaseMask);
+		l1table[a>>20] = L1Section | L1KernelRW | (a&L1SectBaseMask) |
+				L1Cached | L1Buffered;
 
-	/* map the uart so that we can continue using iprint */
-//	uart3regs = mapspecial(UART3REGS, 64);
+	/* map first page of DRAM also into 0xFFFF0000 for the interrupt vectors */
+	t = xspanalloc(BY2PG, 16*1024, 0);
+	memset(t, 0, BY2PG);
+	l1table[0xFFFF0000>>20] = L1PageTable | L1Domain0 | (((ulong)t) & L1PTBaseMask);
+	t[0xF0000>>PGSHIFT] = L2SmallPage | L2KernelRW | PHYSDRAM0;
 
 	/* set up the domain register to cause all domains to obey pte access bits */
 	iprint("setting up domain access\n");
-	putdac(0x55555555);
+	putdac(0xFFFFFFFF);
 
 	/* point to map */
 	iprint("setting tlb map %lux\n", (ulong)l1table);
 	putttb((ulong)l1table);
+
+	/* map the uart so that we can continue using iprint */
+	uart3regs = (Uartregs*)mapspecial(UART3REGS, 64);
+
+	/* enable mmu, and make 0xFFFF0000 the virtual address of the exception vecs */
+	mmuenable();
+
+	iprint("uart3regs now at %lux\n", uart3regs);
 }
 
 /*
@@ -121,26 +128,35 @@ mapspecial(ulong pa, int len)
 	}
 	off = pa - base;
 
-	for(va = REGZERO; va < REGTOP && base >= end; va += OneMeg){
-		if((l1table[va>>20] & L1TypeMask) != L1PageTable){
-
+	for(va = REGZERO; va < REGTOP && base <= end; va += OneMeg){
+		switch(l1table[va>>20] & L1TypeMask){
+		default:
 			/* found unused entry on level 1 table */
 			if(livelarge){
 				if(rv == nil)
-					rv = (ulong*)(va+i*BY2PG+off);
+					rv = (ulong*)(va+off);
 				l1table[va>>20] = L1Section | L1KernelRW |
-							(base&L1SectBaseMask);
+							(base & L1SectBaseMask);
 				base += OneMeg;
 				continue;
-			}
+			} else {
 
-			/* create a page table and keep going */
-			t = xspanalloc(BY2PG, 1024, 0);
-			memset(t, 0, BY2PG);
-			l1table[va>>20] = L1PageTable | L1Domain0 |
-						(((ulong)t) & L1PTBaseMask);
+				/* create an L2 page table and keep going */
+				t = xspanalloc(BY2PG, 1024, 0);
+				memset(t, 0, BY2PG);
+				l1table[va>>20] = L1PageTable | L1Domain0 |
+							(((ulong)t) & L1PTBaseMask);
+			}
+			break;
+		case L1Section:
+			continue;
+		case L1PageTable:
+			if(livelarge)
+				continue;
+			break;
 		}
 
+		/* here if we're using page maps instead of sections */
 		t = (ulong*)(l1table[va>>20] & L1PTBaseMask);
 		for(i = 0; i < OneMeg; i += BY2PG){
 			entry = t[i>>PGSHIFT];
@@ -148,7 +164,7 @@ mapspecial(ulong pa, int len)
 			/* found unused entry on level 2 table */
 			if((entry & L2TypeMask) != L2SmallPage){
 				if(rv == nil)
-					rv = (ulong*)(va+i*BY2PG+off);
+					rv = (ulong*)(va+i+off);
 				t[i>>PGSHIFT] = L2SmallPage | L2KernelRW | 
 						(base & L2PageBaseMask);
 				base += BY2PG;
diff --git a/bitsy/trap.c b/bitsy/trap.c
index cf2cf967a18f80a0bbf1c64f3d2256a9ab54585d..bc92a93c67ce5ad334c5b5d217b845b9e30e547e 100644
--- a/bitsy/trap.c
+++ b/bitsy/trap.c
@@ -19,8 +19,29 @@ evenaddr(ulong addr)
 	}
 }
 
+/*
+ *  set up for exceptioons
+ */
+void
+trapinit(void)
+{
+	/*
+	 *  exceptionvectors points to a prototype in l.s of the
+	 *  exception vectors that save the regs and then call
+	 *  trap().  The actual vectorrs are double mapped
+	 *  to 0xffff0000 and to KZERO.  We write them via
+	 *  KZERO since a data access to them will cause an
+	 *  exception.
+	 */
+	memmove((void*)KZERO, exceptionvectors, 2*4*8);
+	wbflush();
+}
+
+/*
+ *  here on all exceptions with registers saved
+ */
 void
-exception(void)
+trap(Ureg *ureg)
 {
 }
 
diff --git a/carrera/l.s b/carrera/l.s
index cc6a7dd100b2bdfd62588de12ad5fbc581acd916..8ee69c0ee951dd0dfa1b530372865f9216dd283d 100644
--- a/carrera/l.s
+++ b/carrera/l.s
@@ -28,6 +28,8 @@ TEXT	start(SB), $-4
 	MOVW	$TLBROFF, R1
 	MOVW	R1, M(WIRED)
 
+	/* setfcr(FPPDBL|FPRNR|FPINVAL|FPZDIV|FPOVFL) */
+	/* bit 24 is R4000-specific; underflow goes to zero */
 	MOVW	$((0x1C<<7)|(1<<24)), R1
 	MOVW	R1, FCR31	/* permit only inexact and underflow */
 	NOOP
diff --git a/pc/l.s b/pc/l.s
index c4792ac9eace465b3cdf3cafc47453d86e489836..c83941e7679a707dbe45567fa93eb77445f48f15 100644
--- a/pc/l.s
+++ b/pc/l.s
@@ -390,8 +390,10 @@ TEXT fpinit(SB), $0				/* enable and init */
 	FPON
 	FINIT
 	WAIT
-	PUSHW	$0x033E
-	FLDCW	0(SP)				/* ignore underflow/precision, signal others */
+	/* setfcr(FPPDBL|FPRNR|FPINVAL|FPZDIV|FPOVFL) */
+	/* note that low 6 bits are masks, not enables, on this chip */
+	PUSHW	$0x0232
+	FLDCW	0(SP)
 	POPW	AX
 	WAIT
 	RET