Allow initial device mappings (from pmap_table) to be backed up by BATC.

Use this on luna88k to map the bitmap planes of the frame buffer used by
the driver. 10% speedup under X.

7 files changed, 213 insertions, 62 deletions

diff --git a/sys/arch/luna88k/luna88k/pmap_table.c b/sys/arch/luna88k/luna88k/pmap_table.c
index 696afe6cbe7..75c72d5744c 100644
--- a/sys/arch/luna88k/luna88k/pmap_table.c
+++ b/sys/arch/luna88k/luna88k/pmap_table.c
@@ -1,4 +1,4 @@
-/*	$OpenBSD: pmap_table.c,v 1.8 2013/11/13 20:35:59 miod Exp $	*/
+/*	$OpenBSD: pmap_table.c,v 1.9 2013/11/16 18:45:20 miod Exp $	*/
 
 /* 
  * Mach Operating System
@@ -65,8 +65,8 @@ luna88k_board_table[] = {
 	{ MROM_ADDR,		MROM_SPACE,		R,	CI },
 	{ BMAP_RFCNT,		PAGE_SIZE,		RW,	CI },
 	{ BMAP_BMSEL,		PAGE_SIZE,		RW,	CI },
-	{ BMAP_BMP,		BMAP_BMAP0 - BMAP_BMP,	RW,	CI },
-	{ BMAP_BMAP0,		BMAP_BMAP1 - BMAP_BMAP0, RW,	CI },
+	{ BMAP_BMP,		BMAP_BMAP0 - BMAP_BMP,	RW,	CI, TRUE },
+	{ BMAP_BMAP0,		BMAP_BMAP1 - BMAP_BMAP0, RW,	CI, TRUE },
 #if 0 /* not until console or X11 needs them */
 	{ BMAP_BMAP1,		BMAP_BMAP2 - BMAP_BMAP1, RW,	CI },
 	{ BMAP_BMAP2,		BMAP_BMAP3 - BMAP_BMAP2, RW,	CI },
diff --git a/sys/arch/m88k/include/cmmu.h b/sys/arch/m88k/include/cmmu.h
index c0b41465b4f..65a2aa2b19c 100644
--- a/sys/arch/m88k/include/cmmu.h
+++ b/sys/arch/m88k/include/cmmu.h
@@ -1,4 +1,4 @@
-/*	$OpenBSD: cmmu.h,v 1.31 2013/11/02 23:10:30 miod Exp $ */
+/*	$OpenBSD: cmmu.h,v 1.32 2013/11/16 18:45:20 miod Exp $ */
 /*
  * Mach Operating System
  * Copyright (c) 1993-1992 Carnegie Mellon University
@@ -33,6 +33,8 @@
  */
 #if defined(_KERNEL) && !defined(_LOCORE)
 
+#include <machine/mmu.h>
+
 /* machine dependent cmmu function pointer structure */
 struct cmmu_p {
 	cpuid_t (*init)(void);
@@ -107,6 +109,12 @@ extern __cpu_simple_lock_t cmmu_cpu_lock;
 #define DMA_CACHE_SYNC_INVAL	0x01
 #define DMA_CACHE_SYNC		0x02
 
+/*
+ * Current BATC values.
+ */
+
+extern batc_t global_dbatc[BATC_MAX], global_ibatc[BATC_MAX];
+
 #endif	/* _KERNEL && !_LOCORE */
 
 #endif	/* _M88K_CMMU_H_ */
diff --git a/sys/arch/m88k/include/mmu.h b/sys/arch/m88k/include/mmu.h
index fddbed446c1..14cdb63a5e4 100644
--- a/sys/arch/m88k/include/mmu.h
+++ b/sys/arch/m88k/include/mmu.h
@@ -1,4 +1,4 @@
-/*	$OpenBSD: mmu.h,v 1.14 2013/11/02 23:10:30 miod Exp $ */
+/*	$OpenBSD: mmu.h,v 1.15 2013/11/16 18:45:20 miod Exp $ */
 
 /*
  * This file bears almost no resemblance to the original m68k file,
@@ -76,7 +76,7 @@
  * Area descriptors
  */
 
-typedef	u_int32_t	apr_t;
+typedef	uint32_t	apr_t;
 
 #define	APR_V		0x00000001	/* valid bit */
 
@@ -87,21 +87,10 @@ typedef	u_int32_t	apr_t;
 #define PATC_ENTRIES	56
 
 /*
- * BATC entries
- */
-
-#define	BATC_V		0x00000001
-#define	BATC_PROT	0x00000002
-#define	BATC_INH	0x00000004
-#define	BATC_GLOBAL	0x00000008
-#define	BATC_WT		0x00000010
-#define	BATC_SO		0x00000020
-
-/*
  * Segment table entries
  */
 
-typedef u_int32_t	sdt_entry_t;
+typedef uint32_t	sdt_entry_t;
 
 #define	SG_V		0x00000001
 #define	SG_NV		0x00000000
@@ -118,7 +107,7 @@ typedef u_int32_t	sdt_entry_t;
  * Page table entries
  */
 
-typedef u_int32_t	pt_entry_t;
+typedef uint32_t	pt_entry_t;
 
 #define	PG_V		0x00000001
 #define	PG_NV		0x00000000
@@ -165,9 +154,18 @@ typedef u_int32_t	pt_entry_t;
 #define	PDTIDX(va)	(((va) & PDT_MASK) >> PDT_SHIFT)
 
 /*
- * Parameters and macros for BATC
+ * BATC entries
  */
 
+#define	BATC_V		0x00000001
+#define	BATC_PROT	0x00000002
+#define	BATC_INH	0x00000004
+#define	BATC_GLOBAL	0x00000008
+#define	BATC_WT		0x00000010
+#define	BATC_SO		0x00000020
+
+typedef uint32_t	batc_t;
+
 /* 8820x fixed size BATC */
 #define	BATC_BLKSHIFT	19
 #define	BATC_BLKBYTES	(1 << BATC_BLKSHIFT)
diff --git a/sys/arch/m88k/include/pmap_table.h b/sys/arch/m88k/include/pmap_table.h
index b44097ae7b2..b3a987d4633 100644
--- a/sys/arch/m88k/include/pmap_table.h
+++ b/sys/arch/m88k/include/pmap_table.h
@@ -1,4 +1,4 @@
-/*	$OpenBSD: pmap_table.h,v 1.4 2011/03/23 16:54:35 pirofti Exp $	*/
+/*	$OpenBSD: pmap_table.h,v 1.5 2013/11/16 18:45:20 miod Exp $	*/
 /*
  * Mach Operating System
  * Copyright (c) 1992 Carnegie Mellon University
@@ -38,6 +38,7 @@ struct pmap_table {
 	psize_t		size;
 	vm_prot_t	prot;
 	unsigned int	cacheability;
+	boolean_t	may_use_batc;
 };
 
 const struct pmap_table *pmap_table_build(void);
diff --git a/sys/arch/m88k/m88k/m8820x_machdep.c b/sys/arch/m88k/m88k/m8820x_machdep.c
index 90cf25c5480..32bec51eae8 100644
--- a/sys/arch/m88k/m88k/m8820x_machdep.c
+++ b/sys/arch/m88k/m88k/m8820x_machdep.c
@@ -1,4 +1,4 @@
-/*	$OpenBSD: m8820x_machdep.c,v 1.60 2013/11/03 09:42:55 miod Exp $	*/
+/*	$OpenBSD: m8820x_machdep.c,v 1.61 2013/11/16 18:45:20 miod Exp $	*/
 /*
  * Copyright (c) 2004, 2007, 2010, 2011, 2013, Miodrag Vallat.
  *
@@ -173,8 +173,10 @@ void	m8820x_cmmu_inv_locked(int, paddr_t, psize_t);
 void	m8820x_enable_other_cmmu_cache(void);
 #endif
 
-void	m8820x_dbatc_set(cpuid_t, uint, uint32_t);
-void	m8820x_ibatc_set(cpuid_t, uint, uint32_t);
+static inline
+void	m8820x_dbatc_set(cpuid_t, uint, batc_t);
+static inline
+void	m8820x_ibatc_set(cpuid_t, uint, batc_t);
 
 /* Flags passed to m8820x_cmmu_set_*() */
 #define MODE_VAL		0x01
@@ -345,14 +347,16 @@ m8820x_cmmu_wait(int cpu)
  * BATC routines
  */
 
+static inline
 void
-m8820x_dbatc_set(cpuid_t cpu, uint batcno, uint32_t batc)
+m8820x_dbatc_set(cpuid_t cpu, uint batcno, batc_t batc)
 {
 	m8820x_cmmu_set_reg_if_mode(CMMU_BWP(batcno), batc, cpu, DATA_CMMU);
 }
 
+static inline
 void
-m8820x_ibatc_set(cpuid_t cpu, uint batcno, uint32_t batc)
+m8820x_ibatc_set(cpuid_t cpu, uint batcno, batc_t batc)
 {
 	m8820x_cmmu_set_reg_if_mode(CMMU_BWP(batcno), batc, cpu, INST_CMMU);
 }
@@ -362,7 +366,7 @@ m8820x_batc_setup(cpuid_t cpu, apr_t cmode)
 {
 	paddr_t s_text, e_text, s_data, e_data,	e_rodata;
 	uint batcno;
-	uint32_t batc, proto;
+	batc_t batc, proto;
 	extern caddr_t kernelstart;
 	extern caddr_t etext;
 	extern caddr_t erodata;
@@ -392,7 +396,7 @@ m8820x_batc_setup(cpuid_t cpu, apr_t cmode)
 #ifdef DEBUG
 		printf("cpu%d ibat%d %p(%08x)\n", cpu, batcno, s_text, batc);
 #endif
-		m8820x_ibatc_set(cpu, batcno, batc);
+		global_ibatc[batcno] = batc;
 		s_text += BATC_BLKBYTES;
 		if (++batcno == BATC_MAX)
 			break;
@@ -415,11 +419,16 @@ m8820x_batc_setup(cpuid_t cpu, apr_t cmode)
 #ifdef DEBUG
 		printf("cpu%d dbat%d %p(%08x)\n", cpu, batcno, s_data, batc);
 #endif
-		m8820x_dbatc_set(cpu, batcno, batc);
+		global_dbatc[batcno] = batc;
 		s_data += BATC_BLKBYTES;
 		if (++batcno == BATC_MAX)
 			break;
 	}
+
+	for (batcno = 0; batcno < BATC_MAX; batcno++) {
+		m8820x_dbatc_set(cpu, batcno, global_dbatc[batcno]);
+		m8820x_ibatc_set(cpu, batcno, global_ibatc[batcno]);
+	}
 }
 
 /*
diff --git a/sys/arch/m88k/m88k/pmap.c b/sys/arch/m88k/m88k/pmap.c
index a9a73e87ebb..dde3272a098 100644
--- a/sys/arch/m88k/m88k/pmap.c
+++ b/sys/arch/m88k/m88k/pmap.c
@@ -1,4 +1,4 @@
-/*	$OpenBSD: pmap.c,v 1.74 2013/11/03 09:42:55 miod Exp $	*/
+/*	$OpenBSD: pmap.c,v 1.75 2013/11/16 18:45:20 miod Exp $	*/
 
 /*
  * Copyright (c) 2001-2004, 2010, Miodrag Vallat.
@@ -148,13 +148,20 @@ paddr_t	s_firmware;
 psize_t	l_firmware;
 
 /*
+ * Current BATC values.
+ */
+
+batc_t global_dbatc[BATC_MAX];
+batc_t global_ibatc[BATC_MAX];
+
+/*
  * Internal routines
  */
 void		 pmap_changebit(struct vm_page *, int, int);
 void		 pmap_clean_page(paddr_t);
 pt_entry_t	*pmap_expand(pmap_t, vaddr_t, int);
 pt_entry_t	*pmap_expand_kmap(vaddr_t, int);
-void		 pmap_map(paddr_t, psize_t, vm_prot_t, u_int);
+void		 pmap_map(paddr_t, psize_t, vm_prot_t, u_int, boolean_t);
 pt_entry_t	*pmap_pte(pmap_t, vaddr_t);
 void		 pmap_remove_page(struct vm_page *);
 void		 pmap_remove_pte(pmap_t, vaddr_t, pt_entry_t *,
@@ -242,29 +249,91 @@ int
 pmap_translation_info(pmap_t pmap, vaddr_t va, paddr_t *pap, uint32_t *ti)
 {
 	pt_entry_t *pte;
+	vaddr_t var;
+	uint batcno;
 	int s;
 	int rv;
 
 	/*
 	 * Check for a BATC translation first.
-	 * Even though we do not use BATC yet, 88100-based designs (with
-	 * 8820x CMMUs) have two hardwired BATC entries which map the
-	 * upper 1MB (so-called `utility space') 1:1 in supervisor space.
+	 * We only use BATC for supervisor mappings (i.e. pmap_kernel()).
 	 */
+
+	if (pmap == pmap_kernel()) {
+		/*
+		 * 88100-based designs (with 8820x CMMUs) have two hardwired
+		 * BATC entries which map the upper 1MB (so-called
+		 * `utility space') 1:1 in supervisor space.
+		 */
 #ifdef M88100
-	if (CPU_IS88100 && pmap == pmap_kernel()) {
-		if (va >= BATC9_VA) {
-			*pap = va;
-			*ti = BATC9 & CACHE_MASK;
-			return PTI_BATC;
+		if (CPU_IS88100) {
+			if (va >= BATC9_VA) {
+				*pap = va;
+				*ti = 0;
+				if (BATC9 & BATC_INH)
+					*ti |= CACHE_INH;
+				if (BATC9 & BATC_GLOBAL)
+					*ti |= CACHE_GLOBAL;
+				if (BATC9 & BATC_WT)
+					*ti |= CACHE_WT;
+				return PTI_BATC;
+			}
+			if (va >= BATC8_VA) {
+				*pap = va;
+				*ti = 0;
+				if (BATC8 & BATC_INH)
+					*ti |= CACHE_INH;
+				if (BATC8 & BATC_GLOBAL)
+					*ti |= CACHE_GLOBAL;
+				if (BATC8 & BATC_WT)
+					*ti |= CACHE_WT;
+				return PTI_BATC;
+			}
 		}
-		if (va >= BATC8_VA) {
-			*pap = va;
-			*ti = BATC8 & CACHE_MASK;
-			return PTI_BATC;
+#endif
+
+		/*
+		 * Now try all DBATC entries.
+		 * Note that pmap_translation_info() might be invoked (via
+		 * pmap_extract() ) for instruction faults; we *rely* upon
+		 * the fact that all executable mappings covered by IBATC
+		 * will be:
+		 * - read-only, with no RO->RW upgrade allowed
+		 * - dual mapped by ptes, so that pmap_extract() can still
+		 *   return a meaningful result.
+		 * Should this ever change, some kernel interfaces will need
+		 * to be made aware of (and carry on to callees) whether the
+		 * address should be resolved as an instruction or data
+		 * address.
+		 */
+		var = trunc_batc(va);
+		for (batcno = 0; batcno < BATC_MAX; batcno++) {
+			vaddr_t batcva;
+			paddr_t batcpa;
+			batc_t batc;
+
+			batc = global_dbatc[batcno];
+			if ((batc & BATC_V) == 0)
+				continue;
+
+			batcva = (batc << (BATC_BLKSHIFT - BATC_VSHIFT)) &
+			    ~BATC_BLKMASK;
+			if (batcva == var) {
+				batcpa = (batc <<
+				    (BATC_BLKSHIFT - BATC_PSHIFT)) &
+				    ~BATC_BLKMASK;
+				*pap = batcpa + (va - var);
+				*ti = 0;
+				if (batc & BATC_INH)
+					*ti |= CACHE_INH;
+				if (batc & BATC_GLOBAL)
+					*ti |= CACHE_GLOBAL;
+				if (batc & BATC_WT)
+					*ti |= CACHE_WT;
+				return PTI_BATC;
+			}
 		}
 	}
-#endif
 
 	/*
 	 * Check for a regular PTE translation.
@@ -500,11 +569,19 @@ pmap_steal_memory(vsize_t size, vaddr_t *vstartp, vaddr_t *vendp)
  * [INTERNAL]
  * Setup a wired mapping in pmap_kernel(). Similar to pmap_kenter_pa(),
  * but allows explicit cacheability control.
+ * This is only used at bootstrap time. Mappings may also be backed up
+ * by a BATC entry if requested and possible; but note that the BATC
+ * entries set up here may be overwritten by cmmu_batc_setup() later on
+ * (which is harmless since we are creating proper ptes anyway).
  */
 void
-pmap_map(paddr_t pa, psize_t sz, vm_prot_t prot, u_int cmode)
+pmap_map(paddr_t pa, psize_t sz, vm_prot_t prot, u_int cmode,
+    boolean_t may_use_batc)
 {
 	pt_entry_t *pte, npte;
+	batc_t batc;
+	uint npg, batcno;
+	paddr_t curpa;
 
 	DPRINTF(CD_MAP, ("pmap_map(%p, %p, %x, %x)\n",
 	    pa, sz, prot, cmode));
@@ -514,23 +591,71 @@ pmap_map(paddr_t pa, psize_t sz, vm_prot_t prot, u_int cmode)
 		    pa, pa + sz);
 #endif
 
+	sz = round_page(pa + sz) - trunc_page(pa);
+	pa = trunc_page(pa);
+
 	npte = m88k_protection(prot) | cmode | PG_W | PG_V;
 #ifdef M88110
 	if (CPU_IS88110 && m88k_protection(prot) != PG_RO)
 		npte |= PG_M;
 #endif
 
-	sz = atop(round_page(pa + sz) - trunc_page(pa));
-	pa = trunc_page(pa);
-	while (sz-- != 0) {
-		if ((pte = pmap_pte(pmap_kernel(), pa)) == NULL)
-			pte = pmap_expand_kmap(pa, 0);
+	npg = atop(sz);
+	curpa = pa;
+	while (npg-- != 0) {
+		if ((pte = pmap_pte(pmap_kernel(), curpa)) == NULL)
+			pte = pmap_expand_kmap(curpa, 0);
 
-		*pte = npte | pa;
-		pa += PAGE_SIZE;
+		*pte = npte | curpa;
+		curpa += PAGE_SIZE;
 		pmap_kernel()->pm_stats.resident_count++;
 		pmap_kernel()->pm_stats.wired_count++;
 	}
+
+	if (may_use_batc) {
+		sz = round_batc(pa + sz) - trunc_batc(pa);
+		pa = trunc_batc(pa);
+
+		batc = BATC_SO | BATC_V;
+		if ((prot & VM_PROT_WRITE) == 0)
+			batc |= BATC_PROT;
+		if (cmode & CACHE_INH)
+			batc |= BATC_INH;
+		if (cmode & CACHE_WT)
+			batc |= BATC_WT;
+		batc |= BATC_GLOBAL;	/* XXX 88110 SP */
+
+		for (; sz != 0; sz -= BATC_BLKBYTES, pa += BATC_BLKBYTES) {
+			/* check if an existing BATC covers this area */
+			for (batcno = 0; batcno < BATC_MAX; batcno++) {
+				if ((global_dbatc[batcno] & BATC_V) == 0)
+					continue;
+				curpa = (global_dbatc[batcno] <<
+				    (BATC_BLKSHIFT - BATC_PSHIFT)) &
+				    ~BATC_BLKMASK;
+				if (curpa == pa)
+					break;
+			}
+
+			/*
+			 * If there is a BATC covering this range, reuse it.
+			 * We assume all BATC-possible mappings will use the
+			 * same protection and cacheability settings.
+			 */
+			if (batcno != BATC_MAX)
+				continue;
+
+			/* create a new DBATC if possible */
+			for (batcno = BATC_MAX; batcno != 0; batcno--) {
+				if (global_dbatc[batcno - 1] & BATC_V)
+					continue;
+				global_dbatc[batcno - 1] = batc |
+				    ((pa >> BATC_BLKSHIFT) << BATC_PSHIFT) |
+				    ((pa >> BATC_BLKSHIFT) << BATC_VSHIFT);
+				break;
+			}
+		}
+	}
 }
 
 /*
@@ -666,13 +791,14 @@ pmap_bootstrap(paddr_t s_rom, paddr_t e_rom)
 	if (e_rom != s_rom) {
 		s_firmware = s_rom;
 		l_firmware = e_rom - s_rom;
-		pmap_map(s_firmware, l_firmware, UVM_PROT_RW, CACHE_INH);
+		pmap_map(s_firmware, l_firmware, UVM_PROT_RW, CACHE_INH, FALSE);
 	}
 
 	for (ptable = pmap_table_build(); ptable->size != (vsize_t)-1; ptable++)
 		if (ptable->size != 0)
 			pmap_map(ptable->start, ptable->size,
-			    ptable->prot, ptable->cacheability);
+			    ptable->prot, ptable->cacheability,
+			    ptable->may_use_batc);
 
 	/*
 	 * Adjust cache settings according to the hardware we are running on.
diff --git a/sys/arch/mvme88k/mvme88k/m88110.c b/sys/arch/mvme88k/mvme88k/m88110.c
index 5b6152f553f..ac6a3c09552 100644
--- a/sys/arch/mvme88k/mvme88k/m88110.c
+++ b/sys/arch/mvme88k/mvme88k/m88110.c
@@ -1,4 +1,4 @@
-/*	$OpenBSD: m88110.c,v 1.82 2013/11/03 09:42:55 miod Exp $	*/
+/*	$OpenBSD: m88110.c,v 1.83 2013/11/16 18:45:20 miod Exp $	*/
 
 /*
  * Copyright (c) 2010, 2011, Miodrag Vallat.
@@ -181,23 +181,27 @@ const struct cmmu_p cmmu88410 = {
 size_t mc88410_linesz[2] = { 5, 5 };		/* log2 of L2 cache line size */
 size_t mc88410_cachesz[2] = { 256, 256 };	/* L2 cache size in KB */
 
-void m88110_dbatc_set(uint, uint32_t);
-void m88110_ibatc_set(uint, uint32_t);
+static inline
+void m88110_dbatc_set(uint, batc_t);
+static inline
+void m88110_ibatc_set(uint, batc_t);
 void m88110_patc_clear(void);
 
 void m88110_cmmu_wb_locked(paddr_t, psize_t);
 void m88110_cmmu_wbinv_locked(paddr_t, psize_t);
 void m88110_cmmu_inv_locked(paddr_t, psize_t);
 
+static inline
 void
-m88110_dbatc_set(uint batno, uint32_t val)
+m88110_dbatc_set(uint batno, batc_t val)
 {
 	set_dir(batno);
 	set_dbp(val);
 }
 
+static inline
 void
-m88110_ibatc_set(uint batno, uint32_t val)
+m88110_ibatc_set(uint batno, batc_t val)
 {
 	set_iir(batno);
 	set_ibp(val);
@@ -314,7 +318,7 @@ m88110_batc_setup(cpuid_t cpu, apr_t cmode)
 {
 	paddr_t s_text, e_text, s_data, e_data,	e_rodata;
 	uint batcno;
-	uint32_t batc, proto;
+	batc_t batc, proto;
 	extern caddr_t kernelstart;
 	extern caddr_t etext;
 	extern caddr_t erodata;
@@ -344,7 +348,7 @@ m88110_batc_setup(cpuid_t cpu, apr_t cmode)
 #ifdef DEBUG
 		printf("cpu%d ibat%d %p(%08x)\n", cpu, batcno, s_text, batc);
 #endif
-		m88110_ibatc_set(batcno, batc);
+		global_ibatc[batcno] = batc;
 		s_text += BATC_BLKBYTES;
 		if (++batcno == BATC_MAX)
 			break;
@@ -367,11 +371,16 @@ m88110_batc_setup(cpuid_t cpu, apr_t cmode)
 #ifdef DEBUG
 		printf("cpu%d dbat%d %p(%08x)\n", cpu, batcno, s_data, batc);
 #endif
-		m88110_dbatc_set(batcno, batc);
+		global_dbatc[batcno] = batc;
 		s_data += BATC_BLKBYTES;
 		if (++batcno == BATC_MAX)
 			break;
 	}
+
+	for (batcno = 0; batcno < BATC_MAX; batcno++) {
+		m88110_dbatc_set(batcno, global_dbatc[batcno]);
+		m88110_ibatc_set(batcno, global_ibatc[batcno]);
+	}
 }
 
 cpuid_t