ARCv2: mm: Implement cache region flush operations

These are more efficient than the per-line ops

Signed-off-by: Vineet Gupta <vgupta@synopsys.com>

2 files changed, 74 insertions, 0 deletions

diff --git a/arch/arc/include/asm/cache.h b/arch/arc/include/asm/cache.h
index 5008021..16e4577 100644
--- a/arch/arc/include/asm/cache.h
+++ b/arch/arc/include/asm/cache.h
@@ -62,6 +62,8 @@ extern unsigned long perip_base, perip_end;
 #define ARC_REG_IC_BCR		0x77	/* Build Config reg */
 #define ARC_REG_IC_IVIC		0x10
 #define ARC_REG_IC_CTRL		0x11
+#define ARC_REG_IC_IVIR		0x16
+#define ARC_REG_IC_ENDR		0x17
 #define ARC_REG_IC_IVIL		0x19
 #define ARC_REG_IC_PTAG		0x1E
 #define ARC_REG_IC_PTAG_HI	0x1F
@@ -76,6 +78,8 @@ extern unsigned long perip_base, perip_end;
 #define ARC_REG_DC_IVDL		0x4A
 #define ARC_REG_DC_FLSH		0x4B
 #define ARC_REG_DC_FLDL		0x4C
+#define ARC_REG_DC_STARTR	0x4D
+#define ARC_REG_DC_ENDR		0x4E
 #define ARC_REG_DC_PTAG		0x5C
 #define ARC_REG_DC_PTAG_HI	0x5F
 
@@ -83,6 +87,8 @@ extern unsigned long perip_base, perip_end;
 #define DC_CTRL_DIS		0x001
 #define DC_CTRL_INV_MODE_FLUSH	0x040
 #define DC_CTRL_FLUSH_STATUS	0x100
+#define DC_CTRL_RGN_OP_INV	0x200
+#define DC_CTRL_RGN_OP_MSK	0xE00
 
 /*System-level cache (L2 cache) related Auxiliary registers */
 #define ARC_REG_SLC_CFG		0x901
diff --git a/arch/arc/mm/cache.c b/arch/arc/mm/cache.c
index 18132eb..8401fcb 100644
--- a/arch/arc/mm/cache.c
+++ b/arch/arc/mm/cache.c
@@ -21,6 +21,10 @@
 #include <asm/cachectl.h>
 #include <asm/setup.h>
 
+#ifdef CONFIG_ISA_ARCV2
+#define USE_RGN_FLSH	1
+#endif
+
 static int l2_line_sz;
 static int ioc_exists;
 int slc_enable = 1, ioc_enable = 1;
@@ -332,6 +336,8 @@ void __cache_line_loop_v3(phys_addr_t paddr, unsigned long vaddr,
 	}
 }
 
+#ifndef USE_RGN_FLSH
+
 /*
  * In HS38x (MMU v4), I-cache is VIPT (can alias), D-cache is PIPT
  * Here's how cache ops are implemented
@@ -394,6 +400,68 @@ void __cache_line_loop_v4(phys_addr_t paddr, unsigned long vaddr,
 	}
 }
 
+#else
+
+/*
+ * optimized flush operation which takes a region as opposed to iterating per line
+ */
+static inline
+void __cache_line_loop_v4(phys_addr_t paddr, unsigned long vaddr,
+			  unsigned long sz, const int op, const int full_page)
+{
+	const unsigned int ctl = ARC_REG_DC_CTRL;
+	unsigned int s, e, val;
+
+	/* Only for Non aliasing I-cache in HS38 */
+	if (op == OP_INV_IC) {
+		s = ARC_REG_IC_IVIR;
+		e = ARC_REG_IC_ENDR;
+	} else {
+		s = ARC_REG_DC_STARTR;
+		e = ARC_REG_DC_ENDR;
+	}
+
+	if (!full_page) {
+		/* for any leading gap between @paddr and start of cache line */
+		sz += paddr & ~CACHE_LINE_MASK;
+		paddr &= CACHE_LINE_MASK;
+
+		/*
+		 *  account for any trailing gap to end of cache line
+		 *  this is equivalent to DIV_ROUND_UP() in line ops above
+		 */
+		sz += L1_CACHE_BYTES - 1;
+	}
+
+	if (is_pae40_enabled()) {
+		/* TBD: check if crossing 4TB boundary */
+		if (op == OP_INV_IC)
+			write_aux_reg(ARC_REG_IC_PTAG_HI, (u64)paddr >> 32);
+		else
+			write_aux_reg(ARC_REG_DC_PTAG_HI, (u64)paddr >> 32);
+	}
+
+	/*
+	 * Flush / Invalidate is provided by DC_CTRL.RNG_OP 0 or 1
+	 * Flush-n-invalidate additionally uses setting DC_CTRL.IM = 1
+	 * just as for line ops which is handled in __before_dc_op()
+	 */
+	val = read_aux_reg(ctl) & ~DC_CTRL_RGN_OP_MSK;
+
+	if (op & OP_INV)
+		val |= DC_CTRL_RGN_OP_INV;
+
+	write_aux_reg(ctl, val);
+
+	/* ENDR needs to be set ahead of START */
+	write_aux_reg(e, paddr + sz);	/* ENDR is exclusive */
+	write_aux_reg(s, paddr);
+
+	/* caller waits on DC_CTRL.FS */
+}
+
+#endif
+
 #if (CONFIG_ARC_MMU_VER < 3)
 #define __cache_line_loop	__cache_line_loop_v2
 #elif (CONFIG_ARC_MMU_VER == 3)