arch/arm/mach-mvebu/armada3700/cpu.c - platform/external/u-boot - Gitiles

 // SPDX-License-Identifier: GPL-2.0+
 /*
  * Copyright (C) 2016 Stefan Roese <sr@denx.de>
  * Copyright (C) 2020 Marek Behún <kabel@kernel.org>
  */

 #include <common.h>
 #include <cpu_func.h>
 #include <dm.h>
 #include <fdtdec.h>
 #include <fdt_support.h>
 #include <init.h>
 #include <asm/global_data.h>
 #include <linux/bitops.h>
 #include <linux/libfdt.h>
 #include <linux/sizes.h>
 #include <asm/io.h>
 #include <asm/system.h>
 #include <asm/arch/cpu.h>
 #include <asm/arch/soc.h>
 #include <asm/armv8/mmu.h>
 #include <sort.h>

 /* Armada 3700 */
 #define MVEBU_GPIO_NB_REG_BASE		(MVEBU_REGISTER(0x13800))

 #define MVEBU_TEST_PIN_LATCH_N		(MVEBU_GPIO_NB_REG_BASE + 0x8)
 #define MVEBU_XTAL_MODE_MASK		BIT(9)
 #define MVEBU_XTAL_MODE_OFFS		9
 #define MVEBU_XTAL_CLOCK_25MHZ		0x0
 #define MVEBU_XTAL_CLOCK_40MHZ		0x1

 #define MVEBU_NB_WARM_RST_REG		(MVEBU_GPIO_NB_REG_BASE + 0x40)
 #define MVEBU_NB_WARM_RST_MAGIC_NUM	0x1d1e

 /* Armada 3700 CPU Address Decoder registers */
 #define MVEBU_CPU_DEC_WIN_REG_BASE	(size_t)(MVEBU_REGISTER(0xcf00))
 #define MVEBU_CPU_DEC_WIN_CTRL(w) \
 	(MVEBU_CPU_DEC_WIN_REG_BASE + ((w) << 4))
 #define MVEBU_CPU_DEC_WIN_CTRL_EN	BIT(0)
 #define MVEBU_CPU_DEC_WIN_CTRL_TGT_MASK	0xf
 #define MVEBU_CPU_DEC_WIN_CTRL_TGT_OFFS	4
 #define MVEBU_CPU_DEC_WIN_CTRL_TGT_DRAM	0
 #define MVEBU_CPU_DEC_WIN_CTRL_TGT_PCIE	2
 #define MVEBU_CPU_DEC_WIN_SIZE(w)	(MVEBU_CPU_DEC_WIN_CTRL(w) + 0x4)
 #define MVEBU_CPU_DEC_WIN_BASE(w)	(MVEBU_CPU_DEC_WIN_CTRL(w) + 0x8)
 #define MVEBU_CPU_DEC_WIN_REMAP(w)	(MVEBU_CPU_DEC_WIN_CTRL(w) + 0xc)
 #define MVEBU_CPU_DEC_WIN_GRANULARITY	16
 #define MVEBU_CPU_DEC_WINS		5
 #define MVEBU_CPU_DEC_CCI_BASE		(MVEBU_CPU_DEC_WIN_REG_BASE + 0xe0)
 #define MVEBU_CPU_DEC_ROM_BASE		(MVEBU_CPU_DEC_WIN_REG_BASE + 0xf4)

 #define MAX_MEM_MAP_REGIONS		(MVEBU_CPU_DEC_WINS + 4)

 #define A3700_PTE_BLOCK_NORMAL \
 	(PTE_BLOCK_MEMTYPE(MT_NORMAL) | PTE_BLOCK_INNER_SHARE)
 #define A3700_PTE_BLOCK_DEVICE \
 	(PTE_BLOCK_MEMTYPE(MT_DEVICE_NGNRNE) | PTE_BLOCK_NON_SHARE)

 DECLARE_GLOBAL_DATA_PTR;

 static struct mm_region mvebu_mem_map[MAX_MEM_MAP_REGIONS] = {
 	{
 		/*
 		 * SRAM, MMIO regions
 		 * Don't remove this, build_mem_map needs it.
 		 */
 		.phys = SOC_REGS_PHY_BASE,
 		.virt = SOC_REGS_PHY_BASE,
 		.size = 0x02000000UL,	/* 32MiB internal registers */
 		.attrs = A3700_PTE_BLOCK_DEVICE
 	},
 };

 struct mm_region *mem_map = mvebu_mem_map;

 static int get_cpu_dec_win(int win, u32 *tgt, u32 *base, u32 *size)
 {
 	u32 reg;

 	reg = readl(MVEBU_CPU_DEC_WIN_CTRL(win));
 	if (!(reg & MVEBU_CPU_DEC_WIN_CTRL_EN))
 		return -1;

 	if (tgt) {
 		reg >>= MVEBU_CPU_DEC_WIN_CTRL_TGT_OFFS;
 		reg &= MVEBU_CPU_DEC_WIN_CTRL_TGT_MASK;
 		*tgt = reg;
 	}

 	if (base) {
 		reg = readl(MVEBU_CPU_DEC_WIN_BASE(win));
 		*base = reg << MVEBU_CPU_DEC_WIN_GRANULARITY;
 	}

 	if (size) {
 		/*
 		 * Window size is encoded as the number of 1s from LSB to MSB,
 		 * followed by 0s. The number of 1s specifies the size in 64 KiB
 		 * granularity.
 		 */
 		reg = readl(MVEBU_CPU_DEC_WIN_SIZE(win));
 		*size = ((reg + 1) << MVEBU_CPU_DEC_WIN_GRANULARITY);
 	}

 	return 0;
 }

 /*
  * Builds mem_map according to CPU Address Decoder settings, which were set by
  * the TIMH image on the Cortex-M3 secure processor, or by ARM Trusted Firmware
  */
 static void build_mem_map(void)
 {
 	int win, region;
 	u32 reg;

 	region = 1;

 	/* CCI-400 */
 	reg = readl(MVEBU_CPU_DEC_CCI_BASE);
 	mvebu_mem_map[region].phys = reg << 20;
 	mvebu_mem_map[region].virt = reg << 20;
 	mvebu_mem_map[region].size = SZ_64K;
 	mvebu_mem_map[region].attrs = A3700_PTE_BLOCK_DEVICE;
 	++region;

 	/* AP BootROM */
 	reg = readl(MVEBU_CPU_DEC_ROM_BASE);
 	mvebu_mem_map[region].phys = reg << 20;
 	mvebu_mem_map[region].virt = reg << 20;
 	mvebu_mem_map[region].size = SZ_1M;
 	mvebu_mem_map[region].attrs = A3700_PTE_BLOCK_NORMAL;
 	++region;

 	for (win = 0; win < MVEBU_CPU_DEC_WINS; ++win) {
 		u32 base, tgt, size;
 		u64 attrs;

 		/* skip disabled windows */
 		if (get_cpu_dec_win(win, &tgt, &base, &size))
 			continue;

 		if (tgt == MVEBU_CPU_DEC_WIN_CTRL_TGT_DRAM)
 			attrs = A3700_PTE_BLOCK_NORMAL;
 		else if (tgt == MVEBU_CPU_DEC_WIN_CTRL_TGT_PCIE)
 			attrs = A3700_PTE_BLOCK_DEVICE;
 		else
 			/* skip windows with other targets */
 			continue;

 		mvebu_mem_map[region].phys = base;
 		mvebu_mem_map[region].virt = base;
 		mvebu_mem_map[region].size = size;
 		mvebu_mem_map[region].attrs = attrs;
 		++region;
 	}

 	/* add list terminator */
 	mvebu_mem_map[region].size = 0;
 	mvebu_mem_map[region].attrs = 0;
 }

 void enable_caches(void)
 {
 	icache_enable();
 	dcache_enable();
 }

 int a3700_dram_init(void)
 {
 	int win;

 	build_mem_map();

 	gd->ram_size = 0;
 	for (win = 0; win < MVEBU_CPU_DEC_WINS; ++win) {
 		u32 base, tgt, size;

 		/* skip disabled windows */
 		if (get_cpu_dec_win(win, &tgt, &base, &size))
 			continue;

 		/* skip non-DRAM windows */
 		if (tgt != MVEBU_CPU_DEC_WIN_CTRL_TGT_DRAM)
 			continue;

 		/*
 		 * It is possible that one image was built for boards with
 		 * different RAM sizes, for example 512 MiB and 1 GiB.
 		 * We therefore try to determine the actual RAM size in the
 		 * window with get_ram_size.
 		 */
 		gd->ram_size += get_ram_size((void *)(size_t)base, size);
 	}

 	return 0;
 }

 struct a3700_dram_window {
 	size_t base, size;
 };

 static int dram_win_cmp(const void *a, const void *b)
 {
 	size_t ab, bb;

 	ab = ((const struct a3700_dram_window *)a)->base;
 	bb = ((const struct a3700_dram_window *)b)->base;

 	if (ab < bb)
 		return -1;
 	else if (ab > bb)
 		return 1;
 	else
 		return 0;
 }

 int a3700_dram_init_banksize(void)
 {
 	struct a3700_dram_window dram_wins[MVEBU_CPU_DEC_WINS];
 	int bank, win, ndram_wins;
 	u32 last_end;
 	size_t size;

 	ndram_wins = 0;
 	for (win = 0; win < MVEBU_CPU_DEC_WINS; ++win) {
 		u32 base, tgt, size;

 		/* skip disabled windows */
 		if (get_cpu_dec_win(win, &tgt, &base, &size))
 			continue;

 		/* skip non-DRAM windows */
 		if (tgt != MVEBU_CPU_DEC_WIN_CTRL_TGT_DRAM)
 			continue;

 		dram_wins[win].base = base;
 		dram_wins[win].size = size;
 		++ndram_wins;
 	}

 	qsort(dram_wins, ndram_wins, sizeof(dram_wins[0]), dram_win_cmp);

 	bank = 0;
 	last_end = -1;

 	for (win = 0; win < ndram_wins; ++win) {
 		/* again determining actual RAM size as in a3700_dram_init */
 		size = get_ram_size((void *)dram_wins[win].base,
 				    dram_wins[win].size);

 		/*
 		 * Check if previous window ends as the current starts. If yes,
 		 * merge these windows into one "bank". This is possible by this
 		 * simple check thanks to mem_map regions being qsorted in
 		 * build_mem_map.
 		 */
 		if (last_end == dram_wins[win].base) {
 			gd->bd->bi_dram[bank - 1].size += size;
 			last_end += size;
 		} else {
 			if (bank == CONFIG_NR_DRAM_BANKS) {
 				printf("Need more CONFIG_NR_DRAM_BANKS\n");
 				return -ENOBUFS;
 			}

 			gd->bd->bi_dram[bank].start = dram_wins[win].base;
 			gd->bd->bi_dram[bank].size = size;
 			last_end = dram_wins[win].base + size;
 			++bank;
 		}
 	}

 	/*
 	 * If there is more place for DRAM BANKS definitions than needed, fill
 	 * the rest with zeros.
 	 */
 	for (; bank < CONFIG_NR_DRAM_BANKS; ++bank) {
 		gd->bd->bi_dram[bank].start = 0;
 		gd->bd->bi_dram[bank].size = 0;
 	}

 	return 0;
 }

 static u32 find_pcie_window_base(void)
 {
 	int win;

 	for (win = 0; win < MVEBU_CPU_DEC_WINS; ++win) {
 		u32 base, tgt;

 		/* skip disabled windows */
 		if (get_cpu_dec_win(win, &tgt, &base, NULL))
 			continue;

 		if (tgt == MVEBU_CPU_DEC_WIN_CTRL_TGT_PCIE)
 			return base;
 	}

 	return -1;
 }

 static int fdt_setprop_inplace_u32_partial(void *blob, int node,
 					   const char *name,
 					   u32 idx, u32 val)
 {
 	val = cpu_to_fdt32(val);

 	return fdt_setprop_inplace_namelen_partial(blob, node, name,
 						   strlen(name),
 						   idx * sizeof(u32),
 						   &val, sizeof(u32));
 }

 int a3700_fdt_fix_pcie_regions(void *blob)
 {
 	u32 base, lowest_cpu_addr, fix_offset;
 	int pci_cells, cpu_cells, size_cells;
 	const u32 *ranges;
 	int node, pnode;
 	int ret, i, len;

 	base = find_pcie_window_base();
 	if (base == -1)
 		return -ENOENT;

 	node = fdt_node_offset_by_compatible(blob, -1, "marvell,armada-3700-pcie");
 	if (node < 0)
 		return node;

 	ranges = fdt_getprop(blob, node, "ranges", &len);
 	if (!ranges || !len || len % sizeof(u32))
 		return -EINVAL;

 	/*
 	 * The "ranges" property is an array of
 	 *   { <PCI address> <CPU address> <size in PCI address space> }
 	 * where number of PCI address cells and size cells is stored in the
 	 * "#address-cells" and "#size-cells" properties of the same node
 	 * containing the "ranges" property and number of CPU address cells
 	 * is stored in the parent's "#address-cells" property.
 	 *
 	 * All 3 elements can span a diffent number of cells. Fetch them.
 	 */
 	pnode = fdt_parent_offset(blob, node);
 	pci_cells = fdt_address_cells(blob, node);
 	cpu_cells = fdt_address_cells(blob, pnode);
 	size_cells = fdt_size_cells(blob, node);

 	/* PCI addresses always use 3 cells */
 	if (pci_cells != 3)
 		return -EINVAL;

 	/* CPU addresses on Armada 37xx always use 2 cells */
 	if (cpu_cells != 2)
 		return -EINVAL;

 	for (i = 0; i < len / sizeof(u32);
 	     i += pci_cells + cpu_cells + size_cells) {
 		/*
 		 * Parent CPU addresses on Armada 37xx are always 32-bit, so
 		 * check that the high word is zero.
 		 */
 		if (fdt32_to_cpu(ranges[i + pci_cells]))
 			return -EINVAL;

 		if (i == 0 ||
 		    fdt32_to_cpu(ranges[i + pci_cells + 1]) < lowest_cpu_addr)
 			lowest_cpu_addr = fdt32_to_cpu(ranges[i + pci_cells + 1]);
 	}

 	/* Calculate fixup offset from the lowest (first) CPU address */
 	fix_offset = base - lowest_cpu_addr;

 	/* If fixup offset is zero there is nothing to fix */
 	if (!fix_offset)
 		return 0;

 	/*
 	 * Fix each CPU address and corresponding PCI address if PCI address
 	 * is not already remapped (has the same value)
 	 */
 	for (i = 0; i < len / sizeof(u32);
 	     i += pci_cells + cpu_cells + size_cells) {
 		u32 cpu_addr;
 		u64 pci_addr;
 		int idx;

 		/* Fix CPU address */
 		idx = i + pci_cells + cpu_cells - 1;
 		cpu_addr = fdt32_to_cpu(ranges[idx]);
 		ret = fdt_setprop_inplace_u32_partial(blob, node, "ranges", idx,
 						      cpu_addr + fix_offset);
 		if (ret)
 			return ret;

 		/* Fix PCI address only if it isn't remapped (is same as CPU) */
 		idx = i + pci_cells - 1;
 		pci_addr = ((u64)fdt32_to_cpu(ranges[idx - 1]) << 32) |
 			   fdt32_to_cpu(ranges[idx]);
 		if (cpu_addr != pci_addr)
 			continue;

 		ret = fdt_setprop_inplace_u32_partial(blob, node, "ranges", idx,
 						      cpu_addr + fix_offset);
 		if (ret)
 			return ret;
 	}

 	return 0;
 }

 void reset_cpu(void)
 {
 	/*
 	 * Write magic number of 0x1d1e to North Bridge Warm Reset register
 	 * to trigger warm reset
 	 */
 	writel(MVEBU_NB_WARM_RST_MAGIC_NUM, MVEBU_NB_WARM_RST_REG);
 }

 /*
  * get_ref_clk
  *
  * return: reference clock in MHz (25 or 40)
  */
 u32 get_ref_clk(void)
 {
 	u32 regval;

 	regval = (readl(MVEBU_TEST_PIN_LATCH_N) & MVEBU_XTAL_MODE_MASK) >>
 		MVEBU_XTAL_MODE_OFFS;

 	if (regval == MVEBU_XTAL_CLOCK_25MHZ)
 		return 25;
 	else
 		return 40;
 }
	// SPDX-License-Identifier: GPL-2.0+
	/*
	* Copyright (C) 2016 Stefan Roese <sr@denx.de>
	* Copyright (C) 2020 Marek Behún <kabel@kernel.org>
	*/

	#include <common.h>
	#include <cpu_func.h>
	#include <dm.h>
	#include <fdtdec.h>
	#include <fdt_support.h>
	#include <init.h>
	#include <asm/global_data.h>
	#include <linux/bitops.h>
	#include <linux/libfdt.h>
	#include <linux/sizes.h>
	#include <asm/io.h>
	#include <asm/system.h>
	#include <asm/arch/cpu.h>
	#include <asm/arch/soc.h>
	#include <asm/armv8/mmu.h>
	#include <sort.h>

	/* Armada 3700 */
	#define MVEBU_GPIO_NB_REG_BASE (MVEBU_REGISTER(0x13800))

	#define MVEBU_TEST_PIN_LATCH_N (MVEBU_GPIO_NB_REG_BASE + 0x8)
	#define MVEBU_XTAL_MODE_MASK BIT(9)
	#define MVEBU_XTAL_MODE_OFFS 9
	#define MVEBU_XTAL_CLOCK_25MHZ 0x0
	#define MVEBU_XTAL_CLOCK_40MHZ 0x1

	#define MVEBU_NB_WARM_RST_REG (MVEBU_GPIO_NB_REG_BASE + 0x40)
	#define MVEBU_NB_WARM_RST_MAGIC_NUM 0x1d1e

	/* Armada 3700 CPU Address Decoder registers */
	#define MVEBU_CPU_DEC_WIN_REG_BASE (size_t)(MVEBU_REGISTER(0xcf00))
	#define MVEBU_CPU_DEC_WIN_CTRL(w) \
	(MVEBU_CPU_DEC_WIN_REG_BASE + ((w) << 4))
	#define MVEBU_CPU_DEC_WIN_CTRL_EN BIT(0)
	#define MVEBU_CPU_DEC_WIN_CTRL_TGT_MASK 0xf
	#define MVEBU_CPU_DEC_WIN_CTRL_TGT_OFFS 4
	#define MVEBU_CPU_DEC_WIN_CTRL_TGT_DRAM 0
	#define MVEBU_CPU_DEC_WIN_CTRL_TGT_PCIE 2
	#define MVEBU_CPU_DEC_WIN_SIZE(w) (MVEBU_CPU_DEC_WIN_CTRL(w) + 0x4)
	#define MVEBU_CPU_DEC_WIN_BASE(w) (MVEBU_CPU_DEC_WIN_CTRL(w) + 0x8)
	#define MVEBU_CPU_DEC_WIN_REMAP(w) (MVEBU_CPU_DEC_WIN_CTRL(w) + 0xc)
	#define MVEBU_CPU_DEC_WIN_GRANULARITY 16
	#define MVEBU_CPU_DEC_WINS 5
	#define MVEBU_CPU_DEC_CCI_BASE (MVEBU_CPU_DEC_WIN_REG_BASE + 0xe0)
	#define MVEBU_CPU_DEC_ROM_BASE (MVEBU_CPU_DEC_WIN_REG_BASE + 0xf4)

	#define MAX_MEM_MAP_REGIONS (MVEBU_CPU_DEC_WINS + 4)

	#define A3700_PTE_BLOCK_NORMAL \
	(PTE_BLOCK_MEMTYPE(MT_NORMAL) \| PTE_BLOCK_INNER_SHARE)
	#define A3700_PTE_BLOCK_DEVICE \
	(PTE_BLOCK_MEMTYPE(MT_DEVICE_NGNRNE) \| PTE_BLOCK_NON_SHARE)

	DECLARE_GLOBAL_DATA_PTR;

	static struct mm_region mvebu_mem_map[MAX_MEM_MAP_REGIONS] = {
	{
	/*
	* SRAM, MMIO regions
	* Don't remove this, build_mem_map needs it.
	*/
	.phys = SOC_REGS_PHY_BASE,
	.virt = SOC_REGS_PHY_BASE,
	.size = 0x02000000UL, /* 32MiB internal registers */
	.attrs = A3700_PTE_BLOCK_DEVICE
	},
	};

	struct mm_region *mem_map = mvebu_mem_map;

	static int get_cpu_dec_win(int win, u32 tgt, u32 base, u32 *size)
	{
	u32 reg;

	reg = readl(MVEBU_CPU_DEC_WIN_CTRL(win));
	if (!(reg & MVEBU_CPU_DEC_WIN_CTRL_EN))
	return -1;

	if (tgt) {
	reg >>= MVEBU_CPU_DEC_WIN_CTRL_TGT_OFFS;
	reg &= MVEBU_CPU_DEC_WIN_CTRL_TGT_MASK;
	*tgt = reg;
	}

	if (base) {
	reg = readl(MVEBU_CPU_DEC_WIN_BASE(win));
	*base = reg << MVEBU_CPU_DEC_WIN_GRANULARITY;
	}

	if (size) {
	/*
	* Window size is encoded as the number of 1s from LSB to MSB,
	* followed by 0s. The number of 1s specifies the size in 64 KiB
	* granularity.
	*/
	reg = readl(MVEBU_CPU_DEC_WIN_SIZE(win));
	*size = ((reg + 1) << MVEBU_CPU_DEC_WIN_GRANULARITY);
	}

	return 0;
	}

	/*
	* Builds mem_map according to CPU Address Decoder settings, which were set by
	* the TIMH image on the Cortex-M3 secure processor, or by ARM Trusted Firmware
	*/
	static void build_mem_map(void)
	{
	int win, region;
	u32 reg;

	region = 1;

	/* CCI-400 */
	reg = readl(MVEBU_CPU_DEC_CCI_BASE);
	mvebu_mem_map[region].phys = reg << 20;
	mvebu_mem_map[region].virt = reg << 20;
	mvebu_mem_map[region].size = SZ_64K;
	mvebu_mem_map[region].attrs = A3700_PTE_BLOCK_DEVICE;
	++region;

	/* AP BootROM */
	reg = readl(MVEBU_CPU_DEC_ROM_BASE);
	mvebu_mem_map[region].phys = reg << 20;
	mvebu_mem_map[region].virt = reg << 20;
	mvebu_mem_map[region].size = SZ_1M;
	mvebu_mem_map[region].attrs = A3700_PTE_BLOCK_NORMAL;
	++region;

	for (win = 0; win < MVEBU_CPU_DEC_WINS; ++win) {
	u32 base, tgt, size;
	u64 attrs;

	/* skip disabled windows */
	if (get_cpu_dec_win(win, &tgt, &base, &size))
	continue;

	if (tgt == MVEBU_CPU_DEC_WIN_CTRL_TGT_DRAM)
	attrs = A3700_PTE_BLOCK_NORMAL;
	else if (tgt == MVEBU_CPU_DEC_WIN_CTRL_TGT_PCIE)
	attrs = A3700_PTE_BLOCK_DEVICE;
	else
	/* skip windows with other targets */
	continue;

	mvebu_mem_map[region].phys = base;
	mvebu_mem_map[region].virt = base;
	mvebu_mem_map[region].size = size;
	mvebu_mem_map[region].attrs = attrs;
	++region;
	}

	/* add list terminator */
	mvebu_mem_map[region].size = 0;
	mvebu_mem_map[region].attrs = 0;
	}

	void enable_caches(void)
	{
	icache_enable();
	dcache_enable();
	}

	int a3700_dram_init(void)
	{
	int win;

	build_mem_map();

	gd->ram_size = 0;
	for (win = 0; win < MVEBU_CPU_DEC_WINS; ++win) {
	u32 base, tgt, size;

	/* skip disabled windows */
	if (get_cpu_dec_win(win, &tgt, &base, &size))
	continue;

	/* skip non-DRAM windows */
	if (tgt != MVEBU_CPU_DEC_WIN_CTRL_TGT_DRAM)
	continue;

	/*
	* It is possible that one image was built for boards with
	* different RAM sizes, for example 512 MiB and 1 GiB.
	* We therefore try to determine the actual RAM size in the
	* window with get_ram_size.
	*/
	gd->ram_size += get_ram_size((void *)(size_t)base, size);
	}

	return 0;
	}

	struct a3700_dram_window {
	size_t base, size;
	};

	static int dram_win_cmp(const void a, const void b)
	{
	size_t ab, bb;

	ab = ((const struct a3700_dram_window *)a)->base;
	bb = ((const struct a3700_dram_window *)b)->base;

	if (ab < bb)
	return -1;
	else if (ab > bb)
	return 1;
	else
	return 0;
	}

	int a3700_dram_init_banksize(void)
	{
	struct a3700_dram_window dram_wins[MVEBU_CPU_DEC_WINS];
	int bank, win, ndram_wins;
	u32 last_end;
	size_t size;

	ndram_wins = 0;
	for (win = 0; win < MVEBU_CPU_DEC_WINS; ++win) {
	u32 base, tgt, size;

	/* skip disabled windows */
	if (get_cpu_dec_win(win, &tgt, &base, &size))
	continue;

	/* skip non-DRAM windows */
	if (tgt != MVEBU_CPU_DEC_WIN_CTRL_TGT_DRAM)
	continue;

	dram_wins[win].base = base;
	dram_wins[win].size = size;
	++ndram_wins;
	}

	qsort(dram_wins, ndram_wins, sizeof(dram_wins[0]), dram_win_cmp);

	bank = 0;
	last_end = -1;

	for (win = 0; win < ndram_wins; ++win) {
	/* again determining actual RAM size as in a3700_dram_init */
	size = get_ram_size((void *)dram_wins[win].base,
	dram_wins[win].size);

	/*
	* Check if previous window ends as the current starts. If yes,
	* merge these windows into one "bank". This is possible by this
	* simple check thanks to mem_map regions being qsorted in
	* build_mem_map.
	*/
	if (last_end == dram_wins[win].base) {
	gd->bd->bi_dram[bank - 1].size += size;
	last_end += size;
	} else {
	if (bank == CONFIG_NR_DRAM_BANKS) {
	printf("Need more CONFIG_NR_DRAM_BANKS\n");
	return -ENOBUFS;
	}

	gd->bd->bi_dram[bank].start = dram_wins[win].base;
	gd->bd->bi_dram[bank].size = size;
	last_end = dram_wins[win].base + size;
	++bank;
	}
	}

	/*
	* If there is more place for DRAM BANKS definitions than needed, fill
	* the rest with zeros.
	*/
	for (; bank < CONFIG_NR_DRAM_BANKS; ++bank) {
	gd->bd->bi_dram[bank].start = 0;
	gd->bd->bi_dram[bank].size = 0;
	}

	return 0;
	}

	static u32 find_pcie_window_base(void)
	{
	int win;

	for (win = 0; win < MVEBU_CPU_DEC_WINS; ++win) {
	u32 base, tgt;

	/* skip disabled windows */
	if (get_cpu_dec_win(win, &tgt, &base, NULL))
	continue;

	if (tgt == MVEBU_CPU_DEC_WIN_CTRL_TGT_PCIE)
	return base;
	}

	return -1;
	}

	static int fdt_setprop_inplace_u32_partial(void *blob, int node,
	const char *name,
	u32 idx, u32 val)
	{
	val = cpu_to_fdt32(val);

	return fdt_setprop_inplace_namelen_partial(blob, node, name,
	strlen(name),
	idx * sizeof(u32),
	&val, sizeof(u32));
	}

	int a3700_fdt_fix_pcie_regions(void *blob)
	{
	u32 base, lowest_cpu_addr, fix_offset;
	int pci_cells, cpu_cells, size_cells;
	const u32 *ranges;
	int node, pnode;
	int ret, i, len;

	base = find_pcie_window_base();
	if (base == -1)
	return -ENOENT;

	node = fdt_node_offset_by_compatible(blob, -1, "marvell,armada-3700-pcie");
	if (node < 0)
	return node;

	ranges = fdt_getprop(blob, node, "ranges", &len);
	if (!ranges \|\| !len \|\| len % sizeof(u32))
	return -EINVAL;

	/*
	* The "ranges" property is an array of
	* { <PCI address> <CPU address> <size in PCI address space> }
	* where number of PCI address cells and size cells is stored in the
	* "#address-cells" and "#size-cells" properties of the same node
	* containing the "ranges" property and number of CPU address cells
	* is stored in the parent's "#address-cells" property.
	*
	* All 3 elements can span a diffent number of cells. Fetch them.
	*/
	pnode = fdt_parent_offset(blob, node);
	pci_cells = fdt_address_cells(blob, node);
	cpu_cells = fdt_address_cells(blob, pnode);
	size_cells = fdt_size_cells(blob, node);

	/* PCI addresses always use 3 cells */
	if (pci_cells != 3)
	return -EINVAL;

	/* CPU addresses on Armada 37xx always use 2 cells */
	if (cpu_cells != 2)
	return -EINVAL;

	for (i = 0; i < len / sizeof(u32);
	i += pci_cells + cpu_cells + size_cells) {
	/*
	* Parent CPU addresses on Armada 37xx are always 32-bit, so
	* check that the high word is zero.
	*/
	if (fdt32_to_cpu(ranges[i + pci_cells]))
	return -EINVAL;

	if (i == 0 \|\|
	fdt32_to_cpu(ranges[i + pci_cells + 1]) < lowest_cpu_addr)
	lowest_cpu_addr = fdt32_to_cpu(ranges[i + pci_cells + 1]);
	}

	/* Calculate fixup offset from the lowest (first) CPU address */
	fix_offset = base - lowest_cpu_addr;

	/* If fixup offset is zero there is nothing to fix */
	if (!fix_offset)
	return 0;

	/*
	* Fix each CPU address and corresponding PCI address if PCI address
	* is not already remapped (has the same value)
	*/
	for (i = 0; i < len / sizeof(u32);
	i += pci_cells + cpu_cells + size_cells) {
	u32 cpu_addr;
	u64 pci_addr;
	int idx;

	/* Fix CPU address */
	idx = i + pci_cells + cpu_cells - 1;
	cpu_addr = fdt32_to_cpu(ranges[idx]);
	ret = fdt_setprop_inplace_u32_partial(blob, node, "ranges", idx,
	cpu_addr + fix_offset);
	if (ret)
	return ret;

	/* Fix PCI address only if it isn't remapped (is same as CPU) */
	idx = i + pci_cells - 1;
	pci_addr = ((u64)fdt32_to_cpu(ranges[idx - 1]) << 32) \|
	fdt32_to_cpu(ranges[idx]);
	if (cpu_addr != pci_addr)
	continue;

	ret = fdt_setprop_inplace_u32_partial(blob, node, "ranges", idx,
	cpu_addr + fix_offset);
	if (ret)
	return ret;
	}

	return 0;
	}

	void reset_cpu(void)
	{
	/*
	* Write magic number of 0x1d1e to North Bridge Warm Reset register
	* to trigger warm reset
	*/
	writel(MVEBU_NB_WARM_RST_MAGIC_NUM, MVEBU_NB_WARM_RST_REG);
	}

	/*
	* get_ref_clk
	*
	* return: reference clock in MHz (25 or 40)
	*/
	u32 get_ref_clk(void)
	{
	u32 regval;

	regval = (readl(MVEBU_TEST_PIN_LATCH_N) & MVEBU_XTAL_MODE_MASK) >>
	MVEBU_XTAL_MODE_OFFS;

	if (regval == MVEBU_XTAL_CLOCK_25MHZ)
	return 25;
	else
	return 40;
	}