/*
 * Driver for simulating ssp as spi device for pxa
 *
 * Copyright (c) 2009 Marvell Inc.
 * Lei Wen <leiwen@marvell.com>
 *
 * Licensed under the GPL-2 or later.
 */


#include <common.h>
#include <malloc.h>
#include <spi.h>
#include <watchdog.h>
#include <asm/io.h>
#include <asm/arch/pxa2xx_spi.h>
#include <asm/gpio.h>
#include <asm/arch/pxa_dma.h>
#include <asm/arch/cpu.h>

#define to_pxa_spi_slave(s)	container_of(s, struct pxa_spi_slave, slave)

int cs_continuous;

#ifndef CONFIG_PXA_SPI_ENABLE_AUTO_CS
void spi_cs_activate(struct spi_slave *slave)
{
	struct pxa_spi_slave *pss = to_pxa_spi_slave(slave);
	gpio_set_value(slave->cs, pss->gpio_cs_inverted);
}

void spi_cs_deactivate(struct spi_slave *slave)
{
	struct pxa_spi_slave *pss = to_pxa_spi_slave(slave);
	gpio_set_value(slave->cs, !pss->gpio_cs_inverted);
}
#endif

static int spi_pxa_write(struct pxa_spi_slave *pss, unsigned int bitlen)
{
	int wait_timeout = SSP_FLUSH_NUM;
	int bytes = bitlen / pss->slave.wordlen;

	while (--wait_timeout && !(readl(&pss->spi_reg->sssr) & SSSR_TNF))
		;
	if (!wait_timeout) {
		debug("%s: timeout error\n", __func__);
		return -1;
	}

	while (bytes--) {
		if (pss->tx != NULL) {
			if (pss->slave.wordlen > 16) {
				writel(*(u32 *)pss->tx, &pss->spi_reg->ssdr);
				pss->tx += 4;
			} else if (pss->slave.wordlen > 8) {
				writel(*(u16 *)pss->tx, &pss->spi_reg->ssdr);
				pss->tx += 2;
			} else {
				writel(*(u8 *)pss->tx, &pss->spi_reg->ssdr);
				++pss->tx;
			}
		} else
			writel(0, &pss->spi_reg->ssdr);

		while ((readl(&pss->spi_reg->sssr)&0xF10) && wait_timeout--) {
			nop();
			nop();
		}
		if (!wait_timeout) {
			debug("%s: timeout error\n", __func__);
			return -1;
		}
	}

	return 0;
}

static int spi_pxa_read(struct pxa_spi_slave *pss, unsigned int bitlen)
{
	int wait_timeout = SSP_FLUSH_NUM;
	int bytes = bitlen / pss->slave.wordlen;	

	while (--wait_timeout && !(readl(&pss->spi_reg->sssr) & SSSR_TNF))
		;
	if (!wait_timeout) {
		debug("%s: timeout error\n", __func__);
		return -1;
	}

	while (bytes--) {
		wait_timeout = SSP_FLUSH_NUM;	
		writel(0xff, &pss->spi_reg->ssdr);
		while ((readl(&pss->spi_reg->sssr)&0xF10) && wait_timeout--) {
			nop();
			nop();
		}
		if (!wait_timeout) {
			printf("%s: timeout error, sssr:0x%x. byte:0x%x.\n",
				__func__, readl(&pss->spi_reg->sssr), bytes);
			return -1;
		}

		wait_timeout = SSP_FLUSH_NUM;
		while (--wait_timeout && !(readl(&pss->spi_reg->sssr) & SSSR_RNE))
		;
		if (!wait_timeout) {
			printf("%s: timeout error sssr:0x%x.\n", __func__,
					readl(&pss->spi_reg->sssr));
			return -1;
		}

		if (pss->rx != NULL) {
			if (pss->slave.wordlen > 16) {
				*(u32 *)pss->rx = readl(&pss->spi_reg->ssdr);
				pss->rx += 4;
			} else if (pss->slave.wordlen > 8) {
				*(u16 *)pss->rx = readl(&pss->spi_reg->ssdr);
				pss->rx += 2;
			} else {
				*(u8 *)pss->rx = readl(&pss->spi_reg->ssdr);
				++pss->rx;
			}
		} else
			readl(&pss->spi_reg->ssdr);
	}	

	return 0;
}

static int spi_pxa_flush(struct pxa_spi_slave *pss)
{
	unsigned long limit = SSP_FLUSH_NUM;

	do {
		while (readl(&pss->spi_reg->sssr) & SSSR_RNE)
			readl(&pss->spi_reg->ssdr);
	} while ((readl(&pss->spi_reg->sssr) & SSSR_BSY) && limit--);

	writel(SSSR_ROR, &pss->spi_reg->sssr);

	return limit;
}

struct spi_slave *spi_setup_slave(unsigned int bus, unsigned int cs,
		unsigned int max_hz, unsigned int mode)
{
	struct pxa_spi_slave *pss;

	pss = spi_alloc_slave(struct pxa_spi_slave, bus, cs);
	if (!pss)
		return NULL;

	pss->slave.bus = bus;
#ifndef CONFIG_PXA_SPI_ENABLE_AUTO_CS
	pss->slave.cs = cs;
#endif
	pss->spi_reg = (struct ssp_reg *)CONFIG_SYS_SSP_BASE;
#ifdef CONFIG_SPI_52M_ENABLE
	pss->cr0 = SSCR0_MOTO | SSCR0_DATASIZE(DEFAULT_WORD_LEN)
		 | SSCR0_SSE | SSCR0_52MM;
#else
	pss->cr0 = SSCR0_MOTO | SSCR0_DATASIZE(DEFAULT_WORD_LEN)
		 | SSCR0_SSE;
#endif

	pss->cr1 = SSCR1_TTE | SSCR1_TTELP |
		(SSCR1_RXTRESH(RX_THRESH_DEF) & SSCR1_RFT) |
		(SSCR1_TXTRESH(TX_THRESH_DEF) & SSCR1_TFT);

	pss->cr1 &= ~(SSCR1_SPO | SSCR1_SPH);
	pss->cr1 |= (((mode & SPI_CPHA) != 0) ? SSCR1_SPH : 0)
		| (((mode & SPI_CPOL) != 0) ? SSCR1_SPO : 0);

	pss->int_cr1 = SSCR1_TIE | SSCR1_RIE | SSCR1_TINTE;
	pss->clear_sr = SSSR_ROR | SSSR_TINT;

	if (cs) {
#ifndef CONFIG_PXA_SPI_ENABLE_AUTO_CS
		gpio_direction_output(cs, !pss->gpio_cs_inverted);
		pss->gpio_cs_inverted = mode & SPI_CS_HIGH;
		gpio_set_value(cs, !pss->gpio_cs_inverted);
#endif
	}
	cs_continuous = 0;

	return &pss->slave;
}

void spi_free_slave(struct spi_slave *slave)
{
	struct pxa_spi_slave *pss = to_pxa_spi_slave(slave);
	free(pss);
}
#ifdef CONFIG_PXA_DMA
__attribute__ ((aligned(32))) unsigned int dummy[0x400];
#endif
void spi_init(void)
{
	unsigned int reg;
	int clk = SSP_26M; // default 26M
	if (CONFIG_SYS_SSP_BASE == SSP0_BASE)
		reg = APBC_SSP0_CLK_RST;
	else if (CONFIG_SYS_SSP_BASE == SSP1_BASE)
		reg = APBC_SSP1_CLK_RST;
	else if (CONFIG_SYS_SSP_BASE == SSP2_BASE)
		reg = APBC_SSP2_CLK_RST;
	else {
		printf("Fatal error: unsupportted ssp base: 0x%x.\n", CONFIG_SYS_SSP_BASE);
		return;
	}

#ifdef CONFIG_SPI_52M_ENABLE
	clk = SSP_52M;
#else
	#ifdef CONFIG_SPI_26M_ENABLE
		clk = SSP_26M;
	#else
		#ifdef CONFIG_SPI_13M_ENABLE
			clk = SSP_13M;
		#else
			#ifdef CONFIG_SPI_6P5M_ENABLE
				clk = SSP_6P5M;
			#endif
		#endif
	#endif
#endif

	__raw_writel(APBC_SSP_BCLKEN | APBC_SSP_FNCLKEN |
		(clk << APBC_SSP_FNCLKSEL_SHIFT),  reg);

#ifdef CONFIG_PXA_DMA
	memset(dummy, 0x0, 0x400*4);
#endif
	/* Load default SSP configuration */
	writel(0, CONFIG_SYS_SSP_BASE + SSCR0);
	writel(SSCR1_RXTRESH(RX_THRESH_DEF) |
		SSCR1_TXTRESH(TX_THRESH_DEF), CONFIG_SYS_SSP_BASE + SSCR1);
	writel(SSCR0_MOTO | SSCR0_DATASIZE(DEFAULT_WORD_LEN)
			, CONFIG_SYS_SSP_BASE + SSCR0);
	writel(0, CONFIG_SYS_SSP_BASE + SSTO);
	writel(0, CONFIG_SYS_SSP_BASE + SSPSP);

	__raw_writel(DMA_CLK_AXICLK_EN | DMA_CLK_AXI_RST,
		PMUA_DMA_CLK_RES_CTRL);	/* enable DMA clock */
}

int spi_claim_bus(struct spi_slave *slave)
{
	struct pxa_spi_slave *pss = to_pxa_spi_slave(slave);
	slave->wordlen = 8;	/* set default wordlen */

	if (spi_pxa_flush(pss) == 0)
		return -1;

	return 0;
}

void spi_release_bus(struct spi_slave *slave)
{
}

int spi_xfer(struct spi_slave *slave, unsigned int bitlen, const void *dout,
		void *din, unsigned long flags)
{
	struct pxa_spi_slave *pss = to_pxa_spi_slave(slave);
	int ret = 0;

	if (slave->wordlen < 4 || slave->wordlen > 32) {
		printf("pxa2xx_spi: invalid wordlen %d\n", slave->wordlen);
		return -1;
	}

	if (bitlen % slave->wordlen)
		return -1;

	pss->rx = din;
	pss->tx = dout;

	writel(0, &pss->spi_reg->sscr0);
	pss->cr0 &=  ~(0xf);
	if (slave->wordlen > 16)
		pss->cr0 |= SSCR0_EDSS | (slave->wordlen - 1);
	else
		pss->cr0 |= 0xc00000 | (slave->wordlen - 1);
	writel(pss->cr1, &pss->spi_reg->sscr1);
	writel(TIMEOUT_DEF, &pss->spi_reg->ssto);
	writel(pss->cr0, &pss->spi_reg->sscr0);

	if (flags & SPI_XFER_BEGIN) {
#ifndef CONFIG_PXA_SPI_ENABLE_AUTO_CS
		if (!cs_continuous)
			spi_cs_activate(slave);
#endif
	}

	if (dout != NULL && din != NULL) {
		ret = spi_pxa_write(pss, bitlen - slave->wordlen);
		if(ret)
			return -1;
		
		ret = spi_pxa_read(pss, slave->wordlen);
		if(ret)
			return -1;
	} else if (dout != NULL) {
		ret = spi_pxa_write(pss, bitlen);
		if(ret)
			return -1;
	} else if (din != NULL) {
		ret = spi_pxa_read(pss, bitlen);
		if(ret)
			return -1;
	}

	if (flags & SPI_XFER_END) {
		writel(pss->clear_sr, &pss->spi_reg->sssr);
		writel(0, &pss->spi_reg->ssto);
		writel(0, &pss->spi_reg->sscr0);
#ifndef CONFIG_PXA_SPI_ENABLE_AUTO_CS
		if (!cs_continuous)
			spi_cs_deactivate(slave);
#endif
	}
	return ret;
}
#ifdef CONFIG_PXA_DMA
static int spi_dma_read(void *din, unsigned int len)
{
	int i, init = 0;
	if (!init) {
		if (cpu_is_pxa1826_z3() || cpu_is_pxa1826_a0()) {
			dmac_map_device_to_channel(DMAC_SSP2_RX,
				SSP_RX_CHANNEL);
			dmac_map_device_to_channel(DMAC_SSP2_TX,
				SSP_TX_CHANNEL);
		} else {
			dmac_map_device_to_channel(DMAC_SSP2_RX_Z12,
				SSP_RX_CHANNEL);
			dmac_map_device_to_channel(DMAC_SSP2_TX_Z12,
				SSP_TX_CHANNEL);
		}

		dmac_user_aligment(SSP_RX_CHANNEL);
		dmac_user_aligment(SSP_TX_CHANNEL);
		init = 1;
	}

	pxa_dma_read((unsigned int)din, SSP_SSDR, len, SSP_RX_CHANNEL);
	pxa_dma_write(SSP_SSDR, (unsigned int)dummy, len, SSP_TX_CHANNEL);

	dmac_start_transfer(SSP_RX_CHANNEL);
	dmac_start_transfer(SSP_TX_CHANNEL);

	do {
		for (i = 0; i < 500; i++)
			nop();
		if (dmac_read_dcsr(SSP_TX_CHANNEL) & DCSR_STOPSTATE)
			break;

	} while (1);

	do {
		if (dmac_read_dcsr(SSP_RX_CHANNEL) & DCSR_STOPSTATE)
			break;
	} while (1);


	for (i = 0; i < len; i += 4)
		swab32s(din+i);

	if (!(dmac_read_dcsr(SSP_TX_CHANNEL) & DCSR_STOPSTATE)) {
		printf("Error:DMA write operate timeout.\n");
		return 1;
	}
	if (!(dmac_read_dcsr(SSP_RX_CHANNEL) & DCSR_STOPSTATE)) {
		printf("Error:DMA read operate timeout.\n");
		return 1;
	}

	return 0;
}

static int spi_dma_write(void *dout, unsigned int len)
{
	int i, init = 0;

	for (i = 0; i < len; i += 4)
		swab32s(dout+i);

	flush_dcache_all();
	if (!init) {
		if (cpu_is_pxa1826_z3() || cpu_is_pxa1826_a0())
			dmac_map_device_to_channel(DMAC_SSP2_TX,
				SSP_TX_CHANNEL);
		else
			dmac_map_device_to_channel(DMAC_SSP2_TX_Z12,
				SSP_TX_CHANNEL);

		dmac_user_aligment(SSP_TX_CHANNEL);
		init = 1;
	}

	pxa_dma_write(SSP_SSDR, (unsigned int)dout, len, SSP_TX_CHANNEL);
	dmac_start_transfer(SSP_TX_CHANNEL);

	do {
		for (i = 0; i < 500; i++)
			nop();
		if (dmac_read_dcsr(SSP_TX_CHANNEL) & DCSR_STOPSTATE)
			break;
	} while (1);

	if (!(dmac_read_dcsr(SSP_TX_CHANNEL) & DCSR_STOPSTATE)) {
		printf("Error:DMA write operate timeout.\n");
		return 1;
	}

	return 0;
}

int spi_dma_xfer(struct spi_slave *slave, const u8 *cmd, size_t cmd_len,
			const void *dout, const void *din, size_t data_len)
{
	struct pxa_spi_slave *pss = to_pxa_spi_slave(slave);
	int ret = 0;

#ifndef CONFIG_PXA_SPI_ENABLE_AUTO_CS
	spi_cs_activate(slave);
#endif

	if (cmd) {
		cs_continuous = 1;
		spi_xfer(slave, cmd_len*8, cmd, NULL,
			SPI_XFER_BEGIN | SPI_XFER_END);
	}

	if (dout || din) {
		writel(TIMEOUT_DEF, &pss->spi_reg->ssto);
#ifdef CONFIG_SPI_52M_ENABLE
		writel(SSCR1_TTELP | SSCR1_TTE | SSCR1_RXTRESH(0x3)
			| SSCR1_TRAIL | SSCR1_TSRE | SSCR1_RSRE |
			SSCR1_TXTRESH(0x4), &pss->spi_reg->sscr1);
		writel(SSCR0_52MM | SSCR0_SSE | SSCR0_EDSS |
			SSCR0_DATASIZE(0x10), &pss->spi_reg->sscr0);
#else
		writel(SSCR1_TTELP | SSCR1_TTE | SSCR1_RXTRESH(0x9)
			| SSCR1_TRAIL | SSCR1_TSRE | SSCR1_RSRE |
			SSCR1_TXTRESH(0x4), &pss->spi_reg->sscr1);

		writel(SSCR0_SSE | SSCR0_EDSS |
			SSCR0_DATASIZE(0x10), &pss->spi_reg->sscr0);
#endif
	}

	if (dout)
		ret = spi_dma_write((void *)dout, data_len);
	if (din)
		ret = spi_dma_read((void *)din, data_len);

	while (readl(&pss->spi_reg->sssr) & 0xF10)
		;
	if (dout || din) {
		writel(pss->clear_sr, &pss->spi_reg->sssr);
		clrbits_le32(&pss->spi_reg->sscr1, pss->int_cr1);
		writel(0, &pss->spi_reg->sscr0);
#ifndef CONFIG_PXA_SPI_ENABLE_AUTO_CS
		spi_cs_deactivate(slave);
#endif
	}
	cs_continuous = 0;
	return ret;
}
#endif