ARM: Use LPAE for Virtual Address Translation

This change introduces LPAE for virtual address translation. To enable it, set
ARM_LPAE. Boot slows down about 4ms on Tegra124 with LPAE enabled.

TEST=Booted nyan with and without LPAE. Built nyan_big and daisy.
BUG=None
BRANCH=none
Signed-off-by: Daisuke Nojiri <dnojiri@chromium.org>
Tested-by: Daisuke Nojiri <dnojiri@google.com>

Original-Change-Id: I74aa729b6fe6d243f57123dc792302359c661cad
Original-Reviewed-on: https://chromium-review.googlesource.com/187862
Original-Reviewed-by: Julius Werner <jwerner@chromium.org>
Original-Commit-Queue: Daisuke Nojiri <dnojiri@chromium.org>
Original-Tested-by: Daisuke Nojiri <dnojiri@chromium.org>
(cherry picked from commit 6d8c8b2bbdc70555076081eb3bfaabde7b4a398f)
Signed-off-by: Marc Jones <marc.jones@se-eng.com>

Change-Id: I8980375c14758af35f7d5ec5244be963e5462d8a
Reviewed-on: http://review.coreboot.org/7749
Reviewed-by: Stefan Reinauer <stefan.reinauer@coreboot.org>
Tested-by: build bot (Jenkins)

src/arch/arm/Kconfig

@@ -30,4 +30,8 @@ config CPU_HAS_BOOTBLOCK_INIT
 
 config MAINBOARD_HAS_BOOTBLOCK_INIT
 	bool
-	default n
+	default n
+
+config ARM_LPAE
+	bool "Enable LPAE"
+	default n

src/arch/arm/armv7/mmu.c

@@ -27,6 +27,7 @@
  * SUCH DAMAGE.
  */
 
+#include <assert.h>
 #include <config.h>
 #include <stdlib.h>
 #include <stdint.h>
@@ -37,32 +38,29 @@
 #include <arch/cache.h>
 #include <arch/io.h>
 
-static void *const ttb_buff = (void *)CONFIG_TTB_BUFFER;
+#if CONFIG_ARM_LPAE
+/* See B3.6.2 of ARMv7 Architecture Reference Manual */
+/* TODO: Utilize the contiguous hint flag */
+#define ATTR_BASE (\
+	0ULL << 54 |	/* PN. 0:Not restricted */ \
+	0ULL << 53 |	/* PXN. 0:Not restricted */ \
+	1 << 10 |	/* AF. 1:Accessed. This is to prevent access \
+	 	 	 * fault when accessed for the first time */ \
+	0 << 6 |	/* AP[2:1]. 0b00:full access from PL1 */ \
+	0 << 5 |	/* NS. 0:Output address is in Secure space */ \
+	0 << 1 | 	/* block/table. 0:block entry */ \
+	1 << 0		/* validity. 1:valid */ \
+	)
+#define ATTR_NC		(ATTR_BASE | (MAIR_INDX_NC << 2))
+#define ATTR_WT		(ATTR_BASE | (MAIR_INDX_WT << 2))
+#define ATTR_WB		(ATTR_BASE | (MAIR_INDX_WB << 2))
 
-void mmu_disable_range(unsigned long start_mb, unsigned long size_mb)
+#define BLOCK_SHIFT	21
-{
-	unsigned int i;
-	uint32_t *ttb_entry = ttb_buff;
-	printk(BIOS_DEBUG, "Disabling: 0x%08lx:0x%08lx\n",
-			start_mb*MiB, start_mb*MiB + size_mb*MiB - 1);
-
-	for (i = start_mb; i < start_mb + size_mb; i++)
-		writel(0, &ttb_entry[i]);
-
-	for (i = start_mb; i < start_mb + size_mb; i++) {
-		dccmvac((uintptr_t)&ttb_entry[i]);
-		tlbimvaa(i*MiB);
-	}
-}
-
-void mmu_config_range(unsigned long start_mb, unsigned long size_mb,
-		enum dcache_policy policy)
-{
-	unsigned int i;
-	uint32_t attr;
-	uint32_t *ttb_entry = ttb_buff;
-	const char *str = NULL;
 
+typedef uint64_t pgd_t;
+typedef uint64_t pmd_t;
+static const unsigned int denom = 2;
+#else	/* CONFIG_ARM_LPAE */
 	/*
 	 * Section entry bits:
 	 * 31:20 - section base address
@@ -79,19 +77,83 @@ void mmu_config_range(unsigned long start_mb, unsigned long size_mb,
 	 *     2 - B, 1 for bufferable
 	 *  1: 0 - 0b10 to indicate section entry
 	 */
+#define ATTR_BASE	((3 << 10) | 0x2)
+#define ATTR_NC		(ATTR_BASE | (1 << 4))
+#define ATTR_WT		(ATTR_BASE | (1 << 3))
+#define ATTR_WB		(ATTR_BASE | (1 << 3) | (1 << 2))
+
+#define BLOCK_SHIFT	20
+
+typedef uint32_t pgd_t;
+typedef uint32_t pmd_t;
+static const unsigned int denom = 1;
+#endif	/* CONFIG_ARM_LPAE */
+
+static pmd_t *const ttb_buff = (pmd_t *)CONFIG_TTB_BUFFER;
+
+/*
+ * mask/shift/size for pages and blocks
+ */
+#define PAGE_SHIFT	12
+#define PAGE_SIZE	(1UL << PAGE_SHIFT)
+#define PAGE_MASK	~((1UL << PAGE_SHIFT) - 1)
+#define BLOCK_SIZE	(1UL << BLOCK_SHIFT)
+
+/*
+ * MAIR Index
+ */
+#define MAIR_INDX_NC	0
+#define MAIR_INDX_WT	1
+#define MAIR_INDX_WB	2
+
+static void mmu_flush_page_table_entry_range(
+		unsigned long start_mb, unsigned long size_mb)
+{
+	int i;
+
+	/* Flush the page table entries from the dcache. */
+	for (i = start_mb/denom; i*denom < start_mb + size_mb; i++)
+		dccmvac((uintptr_t)&ttb_buff[i]);
+	dsb();
+	/* Invalidate the TLB entries. */
+	for (i = start_mb/denom; i*denom < start_mb + size_mb; i++)
+		tlbimvaa(i*denom*MiB);
+	dsb();
+	isb();
+}
+
+void mmu_disable_range(unsigned long start_mb, unsigned long size_mb)
+{
+	int i;
+
+	printk(BIOS_DEBUG, "Disabling: [0x%08lx:0x%08lx)\n",
+			start_mb*MiB, start_mb*MiB + size_mb*MiB);
+
+	for (i = start_mb/denom; i*denom < start_mb + size_mb; i++)
+		ttb_buff[i] = 0;
+
+	mmu_flush_page_table_entry_range(start_mb, size_mb);
+}
+
+void mmu_config_range(unsigned long start_mb, unsigned long size_mb,
+		enum dcache_policy policy)
+{
+	const char *str = NULL;
+	pmd_t attr;
+	int i;
 
 	switch(policy) {
 	case DCACHE_OFF:
 		/* XN set to avoid prefetches to uncached/unbuffered regions */
-		attr = (0x3 << 10) | (1 << 4) | 0x2;
+		attr = ATTR_NC;
 		str = "off";
 		break;
 	case DCACHE_WRITEBACK:
-		attr =  (0x3 << 10) | (1 << 3) | (1 << 2) | 0x2;
+		attr = ATTR_WB;
 		str = "writeback";
 		break;
 	case DCACHE_WRITETHROUGH:
-		attr =  (0x3 << 10) | (1 << 3) | 0x2;
+		attr = ATTR_WT;
 		str = "writethrough";
 		break;
 	default:
@@ -99,52 +161,86 @@ void mmu_config_range(unsigned long start_mb, unsigned long size_mb,
 		return;
 	}
 
-	printk(BIOS_DEBUG, "Setting dcache policy: 0x%08lx:0x%08lx [%s]\n",
+	printk(BIOS_DEBUG, "Setting dcache policy: [0x%08lx:0x%08lx) [%s]\n",
-			start_mb << 20, ((start_mb + size_mb) << 20) - 1, str);
+			start_mb << 20, ((start_mb + size_mb) << 20), str);
 
 	/* Write out page table entries. */
-	for (i = start_mb; i < start_mb + size_mb; i++)
+	for (i = start_mb/denom; i*denom < start_mb + size_mb; i++)
-		writel((i << 20) | attr, &ttb_entry[i]);
+		ttb_buff[i] = ((pmd_t)i << BLOCK_SHIFT) | attr;
 
-	/* Flush the page table entries from the dcache. */
+	mmu_flush_page_table_entry_range(start_mb, size_mb);
-	for (i = start_mb; i < start_mb + size_mb; i++)
-		dccmvac((uintptr_t)&ttb_entry[i]);
-
-	dsb();
-
-	/* Invalidate the TLB entries. */
-	for (i = start_mb; i < start_mb + size_mb; i++)
-		tlbimvaa(i*MiB);
-
-	dsb();
-	isb();
 }
 
+/*
+ * For coreboot's purposes, we will create a simple identity map.
+ *
+ * If LPAE is disabled, we will create a L1 page
+ * table in RAM with 1MB section translation entries over the 4GB address space.
+ * (ref: section 10.2 and example 15-4 in Cortex-A series programmer's guide)
+ *
+ * If LPAE is enabled, we do two level translation with one L1 table with 4
+ * entries, each covering a 1GB space, and four L2 tables with 512 entries, each
+ * covering a 2MB space.
+ */
 void mmu_init(void)
 {
-	/*
+        if (CONFIG_ARM_LPAE) {
-	 * For coreboot's purposes, we will create a simple L1 page table
+                pgd_t *const pgd_buff = (pgd_t*)(CONFIG_TTB_BUFFER + 16*KiB);
-	 * in RAM with 1MB section translation entries over the 4GB address
+                pmd_t *pmd = ttb_buff;
-	 * space.
+                int i;
-	 * (ref: section 10.2 and example 15-4 in Cortex-A series
+
-	 * programmer's guide)
+                printk(BIOS_DEBUG, "LPAE Translation tables are @ %p\n",
-         */
+                		ttb_buff);
-	printk(BIOS_DEBUG, "Translation table is @ %p\n", ttb_buff);
+                ASSERT((read_mmfr0() & 0xf) >= 5);
+
+                /*
+                 * Set MAIR
+                 * See B4.1.104 of ARMv7 Architecture Reference Manual
+                 */
+                write_mair0(
+                        0x00 << (MAIR_INDX_NC*8) | /* Strongly-ordered,
+                         	 	 	    * Non-Cacheable */
+                        0xaa << (MAIR_INDX_WT*8) | /* Write-Thru,
+                         	 	 	    * Read-Allocate */
+                        0xff << (MAIR_INDX_WB*8)   /* Write-Back,
+                         	 	 	    * Read/Write-Allocate */
+                        );
+
+                /*
+                 * Set up L1 table
+                 * Once set here, L1 table won't be modified by coreboot.
+                 * See B3.6.1 of ARMv7 Architecture Reference Manual
+                 */
+                for (i = 0; i < 4; i++) {
+                        pgd_buff[i] = ((uint32_t)pmd & PAGE_MASK) |
+                                3;	/* 0b11: valid table entry */
+                        pmd += BLOCK_SIZE / PAGE_SIZE;
+                }
+
+                /*
+                 * Set TTBR0
+                 */
+                write_ttbr0((uintptr_t)pgd_buff);
+        } else {
+                printk(BIOS_DEBUG, "Translation table is @ %p\n", ttb_buff);
+
+                /*
+                 * Translation table base 0 address is in bits 31:14-N, where N
+                 * is given by bits 2:0 in TTBCR (which we set to 0). All lower
+                 * bits in this register should be zero for coreboot.
+                 */
+                write_ttbr0((uintptr_t)ttb_buff);
+        }
 
 	/*
-	 * Disable TTBR1 by setting TTBCR.N to 0b000, which means the TTBR0
+	 * Set TTBCR
-	 * table size is 16KB and has indices VA[31:20].
+	 * See B4.1.153 of ARMv7 Architecture Reference Manual
-	 *
+	 * See B3.5.4 and B3.6.4 for how TTBR0 or TTBR1 is selected.
-	 * ref: Arch Ref. Manual for ARMv7-A, B3.5.4,
 	 */
-	write_ttbcr(read_ttbcr() & ~0x3);
+	write_ttbcr(
-
+		CONFIG_ARM_LPAE << 31 |	/* EAE. 1:Enable LPAE */
-	/*
+		0 << 16 | 0 << 0	/* Use TTBR0 for all addresses */
-	 * Translation table base 0 address is in bits 31:14-N, where N is given
+		);
-	 * by bits 2:0 in TTBCR (which we set to 0). All lower bits in this
-	 * register should be zero for coreboot.
-	 */
-	write_ttbr0((uintptr_t)ttb_buff);
 
 	/* disable domain-level checking of permissions */
 	write_dacr(~0);

src/arch/arm/include/armv7/arch/cache.h

@@ -111,10 +111,34 @@ static inline void write_dacr(uint32_t val)
 	asm volatile ("mcr p15, 0, %0, c3, c0, 0" : : "r" (val));
 }
 
+/* read memory model feature register 0 (MMFR0) */
+static inline uint32_t read_mmfr0(void)
+{
+	uint32_t mmfr;
+	asm volatile ("mrc p15, 0, %0, c0, c1, 4" : "=r" (mmfr));
+	return mmfr;
+}
+/* read MAIR0 (memory address indirection register 0) */
+static inline uint32_t read_mair0(void)
+{
+	uint32_t mair;
+	asm volatile ("mrc p15, 0, %0, c10, c2, 0" : "=r" (mair));
+	return mair;
+}
+/* write MAIR0 (memory address indirection register 0) */
+static inline void write_mair0(uint32_t val)
+{
+	asm volatile ("mcr p15, 0, %0, c10, c2, 0" : : "r" (val));
+}
 /* write translation table base register 0 (TTBR0) */
 static inline void write_ttbr0(uint32_t val)
 {
+#if CONFIG_ARM_LPAE
+	asm volatile ("mcrr p15, 0, %[val], %[zero], c2" : :
+			[val] "r" (val), [zero] "r" (0));
+#else
 	asm volatile ("mcr p15, 0, %0, c2, c0, 0" : : "r" (val) : "memory");
+#endif
 }
 
 /* read translation table base control register (TTBCR) */

src/soc/nvidia/tegra124/Kconfig

@@ -1,4 +1,6 @@
 config SOC_NVIDIA_TEGRA124
+	bool
+	default n
 	select ARCH_BOOTBLOCK_ARMV4
 	select ARCH_ROMSTAGE_ARMV7
 	select ARCH_RAMSTAGE_ARMV7
@@ -6,8 +8,8 @@ config SOC_NVIDIA_TEGRA124
 	select BOOTBLOCK_CONSOLE
 	select DYNAMIC_CBMEM
 	select ARM_BOOTBLOCK_CUSTOM
-	bool
+	select ARM_LPAE
-	default n
+
 
 if SOC_NVIDIA_TEGRA124
 
@@ -31,8 +33,7 @@ config BOOTBLOCK_CPU_INIT
 #  handoff that area may be reclaimed for other uses, e.g. CBFS cache.)
 #
 # 0x4000_0000 TTB (16K+32B). 32B is for L1 table of LPAE.
-# 0x4000_4020 CBMEM console area (8K-32B)
+# 0x4000_4020 CBFS mapping cache (96K-32B)
-# 0x4000_6000 CBFS mapping cache (88K)
 # 0x4001_C000 Stack (16KB... don't reduce without comparing LZMA scratchpad!).
 # 0x4002_0000 Bootblock (max 48KB).
 # 0x4002_C000 ROM stage (max 80KB).
@@ -86,15 +87,11 @@ config TTB_BUFFER
 
 config CBFS_CACHE_ADDRESS
 	hex "memory address to put CBFS cache data"
-	default 0x40006000
+	default 0x40004020
 
 config CBFS_CACHE_SIZE
 	hex "size of CBFS cache data"
-	default 0x00016000
+	default 0x00017fe0
-
-config CBMEM_CONSOLE_PRERAM_BASE
-	hex "memory address of the CBMEM console buffer"
-	default 0x40004020
 
 config TEGRA124_MODEL_TD570D
 	bool "TD570D"