Since our caches are snooping, we only need to broadcast cache invalidates

on 88110 designs. Brings a ~8% speedup on GENERIC.MP on 197DP.

3 files changed, 29 insertions, 27 deletions

diff --git a/sys/arch/m88k/include/cmmu.h b/sys/arch/m88k/include/cmmu.h
index 30968c4b944..62858f66fd7 100644
--- a/sys/arch/m88k/include/cmmu.h
+++ b/sys/arch/m88k/include/cmmu.h
@@ -1,4 +1,4 @@
-/*	$OpenBSD: cmmu.h,v 1.22 2009/02/16 23:03:31 miod Exp $ */
+/*	$OpenBSD: cmmu.h,v 1.23 2010/06/22 17:42:35 miod Exp $ */
 /*
  * Mach Operating System
  * Copyright (c) 1993-1992 Carnegie Mellon University
@@ -71,26 +71,24 @@ extern __cpu_simple_lock_t cmmu_cpu_lock;
 
 #define cmmu_init			(cmmu->init)
 #define setup_board_config		(cmmu->setup_board_config)
-#define	cpu_configuration_print(a)	(cmmu->cpu_configuration_print)(a)
+#define	cpu_configuration_print(cpu)	(cmmu->cpu_configuration_print)(cpu)
 #define	cmmu_shutdown			(cmmu->shutdown)
 #define	cmmu_cpu_number			(cmmu->cpu_number)
-#define	cmmu_set_sapr(a)		(cmmu->set_sapr)(a)
-#define	cmmu_set_uapr(a)		(cmmu->set_uapr)(a)
-#define	cmmu_flush_tlb(a, b, c, d) 	(cmmu->flush_tlb)(a, b, c, d)
-#define	cmmu_flush_cache(a, b, c)	(cmmu->flush_cache)(a, b, c)
-#define	cmmu_flush_inst_cache(a, b, c)	(cmmu->flush_inst_cache)(a, b, c)
-#define	dma_cachectl(a, b, c)		(cmmu->dma_cachectl)(a, b, c)
-#define	dma_cachectl_local(a, b, c)	(cmmu->dma_cachectl_local)(a, b, c)
-#define	cmmu_initialize_cpu(a)		(cmmu->initialize_cpu)(a)
+#define	cmmu_set_sapr(apr)		(cmmu->set_sapr)(apr)
+#define	cmmu_set_uapr(apr)		(cmmu->set_uapr)(apr)
+#define	cmmu_flush_tlb(cpu, k, va, c) 	(cmmu->flush_tlb)(cpu, k, va, c)
+#define	cmmu_flush_cache(cpu, pa, s)	(cmmu->flush_cache)(cpu, pa, s)
+#define	cmmu_flush_inst_cache(cpu,pa,s)	(cmmu->flush_inst_cache)(cpu, pa, s)
+#define	dma_cachectl(pa, s, op)		(cmmu->dma_cachectl)(pa, s, op)
+#define	dma_cachectl_local(pa, s, op)	(cmmu->dma_cachectl_local)(pa, s, op)
+#define	cmmu_initialize_cpu(cpu)	(cmmu->initialize_cpu)(cpu)
 
 /*
  * dma_cachectl{,_local}() modes
  */
-#define DMA_CACHE_SYNC		0x00
+#define DMA_CACHE_INV		0x00
 #define DMA_CACHE_SYNC_INVAL	0x01
-#define DMA_CACHE_INV		0x02
-
-#define	DMA_CACHE_MASK		0x03
+#define DMA_CACHE_SYNC		0x02
 
 #endif	/* _KERNEL && !_LOCORE */
 
diff --git a/sys/arch/mvme88k/mvme88k/m197_machdep.c b/sys/arch/mvme88k/mvme88k/m197_machdep.c
index 7c3b6e4afa3..518075852d2 100644
--- a/sys/arch/mvme88k/mvme88k/m197_machdep.c
+++ b/sys/arch/mvme88k/mvme88k/m197_machdep.c
@@ -1,4 +1,4 @@
-/*	$OpenBSD: m197_machdep.c,v 1.41 2009/08/30 12:11:35 miod Exp $	*/
+/*	$OpenBSD: m197_machdep.c,v 1.42 2010/06/22 17:42:37 miod Exp $	*/
 
 /*
  * Copyright (c) 2009 Miodrag Vallat.
@@ -615,8 +615,7 @@ m197_ipi_handler(struct trapframe *eframe)
 			cmmu_flush_inst_cache(ci->ci_cpuid, arg1, arg2);
 		}
 		else if (ipi & CI_IPI_DMA_CACHECTL) {
-			dma_cachectl_local(arg1, arg2 & ~DMA_CACHE_MASK,
-			    arg2 & DMA_CACHE_MASK);
+			dma_cachectl_local(arg1, arg2, DMA_CACHE_INV);
 		}
 
 		return 0;
diff --git a/sys/arch/mvme88k/mvme88k/m88110.c b/sys/arch/mvme88k/mvme88k/m88110.c
index acc86352093..c6a9d3b6c0a 100644
--- a/sys/arch/mvme88k/mvme88k/m88110.c
+++ b/sys/arch/mvme88k/mvme88k/m88110.c
@@ -1,4 +1,4 @@
-/*	$OpenBSD: m88110.c,v 1.64 2010/04/17 22:10:13 miod Exp $	*/
+/*	$OpenBSD: m88110.c,v 1.65 2010/06/22 17:42:37 miod Exp $	*/
 /*
  * Copyright (c) 1998 Steve Murphree, Jr.
  * All rights reserved.
@@ -647,6 +647,14 @@ m88110_cmmu_inval_cache(paddr_t pa, psize_t size)
  * High level cache handling functions (used by bus_dma).
  */
 
+#ifdef MULTIPROCESSOR
+void
+m88110_dma_cachectl_local(paddr_t _pa, psize_t _size, int op)
+{
+	/* Obviously nothing to do. */
+}
+#endif
+
 void
 m88110_dma_cachectl(paddr_t _pa, psize_t _size, int op)
 {
@@ -790,14 +798,11 @@ m88410_dma_cachectl(paddr_t _pa, psize_t _size, int op)
 
 	m88410_dma_cachectl_local(pa, size, op);
 #ifdef MULTIPROCESSOR
-	m197_broadcast_complex_ipi(CI_IPI_DMA_CACHECTL, pa, size | op);
+	/*
+	 * Since snooping is enabled, all we need is to propagate invalidate 
+	 * requests if necessary.
+	 */
+	if (op != DMA_CACHE_SYNC)
+		m197_broadcast_complex_ipi(CI_IPI_DMA_CACHECTL, pa, size);
 #endif
 }
-
-#ifdef MULTIPROCESSOR
-void
-m88110_dma_cachectl_local(paddr_t _pa, psize_t _size, int op)
-{
-	/* Obviously nothing to do. */
-}
-#endif