drivers/memory/emif.c

/*
 * EMIF driver
 *
 * Copyright (C) 2012 Texas Instruments, Inc.
 *
 * Aneesh V <aneesh@ti.com>
 * Santosh Shilimkar <santosh.shilimkar@ti.com>
 *
 * This program is free software; you can redistribute it and/or modify
 * it under the terms of the GNU General Public License version 2 as
 * published by the Free Software Foundation.
 */
#include <linux/kernel.h>
#include <linux/reboot.h>
#include <linux/platform_data/emif_plat.h>
#include <linux/io.h>
#include <linux/device.h>
#include <linux/platform_device.h>
#include <linux/interrupt.h>
#include <linux/slab.h>
#include <linux/seq_file.h>
#include <linux/module.h>
#include <linux/list.h>
#include <linux/spinlock.h>
#include <memory/jedec_ddr.h>
#include "emif.h"

/**
 * struct emif_data - Per device static data for driver's use
 * @duplicate:			Whether the DDR devices attached to this EMIF
 *				instance are exactly same as that on EMIF1. In
 *				this case we can save some memory and processing
 * @temperature_level:		Maximum temperature of LPDDR2 devices attached
 *				to this EMIF - read from MR4 register. If there
 *				are two devices attached to this EMIF, this
 *				value is the maximum of the two temperature
 *				levels.
 * @node:			node in the device list
 * @base:			base address of memory-mapped IO registers.
 * @dev:			device pointer.
 * @addressing			table with addressing information from the spec
 * @regs_cache:			An array of 'struct emif_regs' that stores
 *				calculated register values for different
 *				frequencies, to avoid re-calculating them on
 *				each DVFS transition.
 * @curr_regs:			The set of register values used in the last
 *				frequency change (i.e. corresponding to the
 *				frequency in effect at the moment)
 * @plat_data:			Pointer to saved platform data.
 */
struct emif_data {
	u8				duplicate;
	u8				temperature_level;
	u8				lpmode;
	struct list_head		node;
	unsigned long			irq_state;
	void __iomem			*base;
	struct device			*dev;
	const struct lpddr2_addressing	*addressing;
	struct emif_regs		*regs_cache[EMIF_MAX_NUM_FREQUENCIES];
	struct emif_regs		*curr_regs;
	struct emif_platform_data	*plat_data;
};

static struct emif_data *emif1;
static spinlock_t	emif_lock;
static unsigned long	irq_state;
static u32		t_ck; /* DDR clock period in ps */
static LIST_HEAD(device_list);

/*
 * Calculate the period of DDR clock from frequency value
 */
static void set_ddr_clk_period(u32 freq)
{
	/* Divide 10^12 by frequency to get period in ps */
	t_ck = (u32)DIV_ROUND_UP_ULL(1000000000000ull, freq);
}

/*
 * Get the CL from SDRAM_CONFIG register
 */
static u32 get_cl(struct emif_data *emif)
{
	u32		cl;
	void __iomem	*base = emif->base;

	cl = (readl(base + EMIF_SDRAM_CONFIG) & CL_MASK) >> CL_SHIFT;

	return cl;
}

static void set_lpmode(struct emif_data *emif, u8 lpmode)
{
	u32 temp;
	void __iomem *base = emif->base;

	temp = readl(base + EMIF_POWER_MANAGEMENT_CONTROL);
	temp &= ~LP_MODE_MASK;
	temp |= (lpmode << LP_MODE_SHIFT);
	writel(temp, base + EMIF_POWER_MANAGEMENT_CONTROL);
}

static void do_freq_update(void)
{
	struct emif_data *emif;

	/*
	 * Workaround for errata i728: Disable LPMODE during FREQ_UPDATE
	 *
	 * i728 DESCRIPTION:
	 * The EMIF automatically puts the SDRAM into self-refresh mode
	 * after the EMIF has not performed accesses during
	 * EMIF_PWR_MGMT_CTRL[7:4] REG_SR_TIM number of DDR clock cycles
	 * and the EMIF_PWR_MGMT_CTRL[10:8] REG_LP_MODE bit field is set
	 * to 0x2. If during a small window the following three events
	 * occur:
	 * - The SR_TIMING counter expires
	 * - And frequency change is requested
	 * - And OCP access is requested
	 * Then it causes instable clock on the DDR interface.
	 *
	 * WORKAROUND
	 * To avoid the occurrence of the three events, the workaround
	 * is to disable the self-refresh when requesting a frequency
	 * change. Before requesting a frequency change the software must
	 * program EMIF_PWR_MGMT_CTRL[10:8] REG_LP_MODE to 0x0. When the
	 * frequency change has been done, the software can reprogram
	 * EMIF_PWR_MGMT_CTRL[10:8] REG_LP_MODE to 0x2
	 */
	list_for_each_entry(emif, &device_list, node) {
		if (emif->lpmode == EMIF_LP_MODE_SELF_REFRESH)
			set_lpmode(emif, EMIF_LP_MODE_DISABLE);
	}

	/*
	 * TODO: Do FREQ_UPDATE here when an API
	 * is available for this as part of the new
	 * clock framework
	 */

	list_for_each_entry(emif, &device_list, node) {
		if (emif->lpmode == EMIF_LP_MODE_SELF_REFRESH)
			set_lpmode(emif, EMIF_LP_MODE_SELF_REFRESH);
	}
}

/* Find addressing table entry based on the device's type and density */
static const struct lpddr2_addressing *get_addressing_table(
	const struct ddr_device_info *device_info)
{
	u32		index, type, density;

	type = device_info->type;
	density = device_info->density;

	switch (type) {
	case DDR_TYPE_LPDDR2_S4:
		index = density - 1;
		break;
	case DDR_TYPE_LPDDR2_S2:
		switch (density) {
		case DDR_DENSITY_1Gb:
		case DDR_DENSITY_2Gb:
			index = density + 3;
			break;
		default:
			index = density - 1;
		}
		break;
	default:
		return NULL;
	}

	return &lpddr2_jedec_addressing_table[index];
}

/*
 * Find the the right timing table from the array of timing
 * tables of the device using DDR clock frequency
 */
static const struct lpddr2_timings *get_timings_table(struct emif_data *emif,
		u32 freq)
{
	u32				i, min, max, freq_nearest;
	const struct lpddr2_timings	*timings = NULL;
	const struct lpddr2_timings	*timings_arr = emif->plat_data->timings;
	struct				device *dev = emif->dev;

	/* Start with a very high frequency - 1GHz */
	freq_nearest = 1000000000;

	/*
	 * Find the timings table such that:
	 *  1. the frequency range covers the required frequency(safe) AND
	 *  2. the max_freq is closest to the required frequency(optimal)
	 */
	for (i = 0; i < emif->plat_data->timings_arr_size; i++) {
		max = timings_arr[i].max_freq;
		min = timings_arr[i].min_freq;
		if ((freq >= min) && (freq <= max) && (max < freq_nearest)) {
			freq_nearest = max;
			timings = &timings_arr[i];
		}
	}

	if (!timings)
		dev_err(dev, "%s: couldn't find timings for - %dHz\n",
			__func__, freq);

	dev_dbg(dev, "%s: timings table: freq %d, speed bin freq %d\n",
		__func__, freq, freq_nearest);

	return timings;
}

static u32 get_sdram_ref_ctrl_shdw(u32 freq,
		const struct lpddr2_addressing *addressing)
{
	u32 ref_ctrl_shdw = 0, val = 0, freq_khz, t_refi;

	/* Scale down frequency and t_refi to avoid overflow */
	freq_khz = freq / 1000;
	t_refi = addressing->tREFI_ns / 100;

	/*
	 * refresh rate to be set is 'tREFI(in us) * freq in MHz
	 * division by 10000 to account for change in units
	 */
	val = t_refi * freq_khz / 10000;
	ref_ctrl_shdw |= val << REFRESH_RATE_SHIFT;

	return ref_ctrl_shdw;
}

static u32 get_sdram_tim_1_shdw(const struct lpddr2_timings *timings,
		const struct lpddr2_min_tck *min_tck,
		const struct lpddr2_addressing *addressing)
{
	u32 tim1 = 0, val = 0;

	val = max(min_tck->tWTR, DIV_ROUND_UP(timings->tWTR, t_ck)) - 1;
	tim1 |= val << T_WTR_SHIFT;

	if (addressing->num_banks == B8)
		val = DIV_ROUND_UP(timings->tFAW, t_ck*4);
	else
		val = max(min_tck->tRRD, DIV_ROUND_UP(timings->tRRD, t_ck));
	tim1 |= (val - 1) << T_RRD_SHIFT;

	val = DIV_ROUND_UP(timings->tRAS_min + timings->tRPab, t_ck) - 1;
	tim1 |= val << T_RC_SHIFT;

	val = max(min_tck->tRASmin, DIV_ROUND_UP(timings->tRAS_min, t_ck));
	tim1 |= (val - 1) << T_RAS_SHIFT;

	val = max(min_tck->tWR, DIV_ROUND_UP(timings->tWR, t_ck)) - 1;
	tim1 |= val << T_WR_SHIFT;

	val = max(min_tck->tRCD, DIV_ROUND_UP(timings->tRCD, t_ck)) - 1;
	tim1 |= val << T_RCD_SHIFT;

	val = max(min_tck->tRPab, DIV_ROUND_UP(timings->tRPab, t_ck)) - 1;
	tim1 |= val << T_RP_SHIFT;

	return tim1;
}

static u32 get_sdram_tim_1_shdw_derated(const struct lpddr2_timings *timings,
		const struct lpddr2_min_tck *min_tck,
		const struct lpddr2_addressing *addressing)
{
	u32 tim1 = 0, val = 0;

	val = max(min_tck->tWTR, DIV_ROUND_UP(timings->tWTR, t_ck)) - 1;
	tim1 = val << T_WTR_SHIFT;

	/*
	 * tFAW is approximately 4 times tRRD. So add 1875*4 = 7500ps
	 * to tFAW for de-rating
	 */
	if (addressing->num_banks == B8) {
		val = DIV_ROUND_UP(timings->tFAW + 7500, 4 * t_ck) - 1;
	} else {
		val = DIV_ROUND_UP(timings->tRRD + 1875, t_ck);
		val = max(min_tck->tRRD, val) - 1;
	}
	tim1 |= val << T_RRD_SHIFT;

	val = DIV_ROUND_UP(timings->tRAS_min + timings->tRPab + 1875, t_ck);
	tim1 |= (val - 1) << T_RC_SHIFT;

	val = DIV_ROUND_UP(timings->tRAS_min + 1875, t_ck);
	val = max(min_tck->tRASmin, val) - 1;
	tim1 |= val << T_RAS_SHIFT;

	val = max(min_tck->tWR, DIV_ROUND_UP(timings->tWR, t_ck)) - 1;
	tim1 |= val << T_WR_SHIFT;

	val = max(min_tck->tRCD, DIV_ROUND_UP(timings->tRCD + 1875, t_ck));
	tim1 |= (val - 1) << T_RCD_SHIFT;

	val = max(min_tck->tRPab, DIV_ROUND_UP(timings->tRPab + 1875, t_ck));
	tim1 |= (val - 1) << T_RP_SHIFT;

	return tim1;
}

static u32 get_sdram_tim_2_shdw(const struct lpddr2_timings *timings,
		const struct lpddr2_min_tck *min_tck,
		const struct lpddr2_addressing *addressing,
		u32 type)
{
	u32 tim2 = 0, val = 0;

	val = min_tck->tCKE - 1;
	tim2 |= val << T_CKE_SHIFT;

	val = max(min_tck->tRTP, DIV_ROUND_UP(timings->tRTP, t_ck)) - 1;
	tim2 |= val << T_RTP_SHIFT;

	/* tXSNR = tRFCab_ps + 10 ns(tRFCab_ps for LPDDR2). */
	val = DIV_ROUND_UP(addressing->tRFCab_ps + 10000, t_ck) - 1;
	tim2 |= val << T_XSNR_SHIFT;

	/* XSRD same as XSNR for LPDDR2 */
	tim2 |= val << T_XSRD_SHIFT;

	val = max(min_tck->tXP, DIV_ROUND_UP(timings->tXP, t_ck)) - 1;
	tim2 |= val << T_XP_SHIFT;

	return tim2;
}

static u32 get_sdram_tim_3_shdw(const struct lpddr2_timings *timings,
		const struct lpddr2_min_tck *min_tck,
		const struct lpddr2_addressing *addressing,
		u32 type, u32 ip_rev, u32 derated)
{
	u32 tim3 = 0, val = 0, t_dqsck;

	val = timings->tRAS_max_ns / addressing->tREFI_ns - 1;
	val = val > 0xF ? 0xF : val;
	tim3 |= val << T_RAS_MAX_SHIFT;

	val = DIV_ROUND_UP(addressing->tRFCab_ps, t_ck) - 1;
	tim3 |= val << T_RFC_SHIFT;

	t_dqsck = (derated == EMIF_DERATED_TIMINGS) ?
		timings->tDQSCK_max_derated : timings->tDQSCK_max;
	if (ip_rev == EMIF_4D5)
		val = DIV_ROUND_UP(t_dqsck + 1000, t_ck) - 1;
	else
		val = DIV_ROUND_UP(t_dqsck, t_ck) - 1;

	tim3 |= val << T_TDQSCKMAX_SHIFT;

	val = DIV_ROUND_UP(timings->tZQCS, t_ck) - 1;
	tim3 |= val << ZQ_ZQCS_SHIFT;

	val = DIV_ROUND_UP(timings->tCKESR, t_ck);
	val = max(min_tck->tCKESR, val) - 1;
	tim3 |= val << T_CKESR_SHIFT;

	if (ip_rev == EMIF_4D5) {
		tim3 |= (EMIF_T_CSTA - 1) << T_CSTA_SHIFT;

		val = DIV_ROUND_UP(EMIF_T_PDLL_UL, 128) - 1;
		tim3 |= val << T_PDLL_UL_SHIFT;
	}

	return tim3;
}

static u32 get_read_idle_ctrl_shdw(u8 volt_ramp)
{
	u32 idle = 0, val = 0;

	/*
	 * Maximum value in normal conditions and increased frequency
	 * when voltage is ramping
	 */
	if (volt_ramp)
		val = READ_IDLE_INTERVAL_DVFS / t_ck / 64 - 1;
	else
		val = 0x1FF;

	/*
	 * READ_IDLE_CTRL register in EMIF4D has same offset and fields
	 * as DLL_CALIB_CTRL in EMIF4D5, so use the same shifts
	 */
	idle |= val << DLL_CALIB_INTERVAL_SHIFT;
	idle |= EMIF_READ_IDLE_LEN_VAL << ACK_WAIT_SHIFT;

	return idle;
}

static u32 get_dll_calib_ctrl_shdw(u8 volt_ramp)
{
	u32 calib = 0, val = 0;

	if (volt_ramp == DDR_VOLTAGE_RAMPING)
		val = DLL_CALIB_INTERVAL_DVFS / t_ck / 16 - 1;
	else
		val = 0; /* Disabled when voltage is stable */

	calib |= val << DLL_CALIB_INTERVAL_SHIFT;
	calib |= DLL_CALIB_ACK_WAIT_VAL << ACK_WAIT_SHIFT;

	return calib;
}

static u32 get_ddr_phy_ctrl_1_attilaphy_4d(const struct lpddr2_timings *timings,
	u32 freq, u8 RL)
{
	u32 phy = EMIF_DDR_PHY_CTRL_1_BASE_VAL_ATTILAPHY, val = 0;

	val = RL + DIV_ROUND_UP(timings->tDQSCK_max, t_ck) - 1;
	phy |= val << READ_LATENCY_SHIFT_4D;

	if (freq <= 100000000)
		val = EMIF_DLL_SLAVE_DLY_CTRL_100_MHZ_AND_LESS_ATTILAPHY;
	else if (freq <= 200000000)
		val = EMIF_DLL_SLAVE_DLY_CTRL_200_MHZ_ATTILAPHY;
	else
		val = EMIF_DLL_SLAVE_DLY_CTRL_400_MHZ_ATTILAPHY;

	phy |= val << DLL_SLAVE_DLY_CTRL_SHIFT_4D;

	return phy;
}

static u32 get_phy_ctrl_1_intelliphy_4d5(u32 freq, u8 cl)
{
	u32 phy = EMIF_DDR_PHY_CTRL_1_BASE_VAL_INTELLIPHY, half_delay;

	/*
	 * DLL operates at 266 MHz. If DDR frequency is near 266 MHz,
	 * half-delay is not needed else set half-delay
	 */
	if (freq >= 265000000 && freq < 267000000)
		half_delay = 0;
	else
		half_delay = 1;

	phy |= half_delay << DLL_HALF_DELAY_SHIFT_4D5;
	phy |= ((cl + DIV_ROUND_UP(EMIF_PHY_TOTAL_READ_LATENCY_INTELLIPHY_PS,
			t_ck) - 1) << READ_LATENCY_SHIFT_4D5);

	return phy;
}

static u32 get_ext_phy_ctrl_2_intelliphy_4d5(void)
{
	u32 fifo_we_slave_ratio;

	fifo_we_slave_ratio =  DIV_ROUND_CLOSEST(
		EMIF_INTELLI_PHY_DQS_GATE_OPENING_DELAY_PS * 256 , t_ck);

	return fifo_we_slave_ratio | fifo_we_slave_ratio << 11 |
		fifo_we_slave_ratio << 22;
}

static u32 get_ext_phy_ctrl_3_intelliphy_4d5(void)
{
	u32 fifo_we_slave_ratio;

	fifo_we_slave_ratio =  DIV_ROUND_CLOSEST(
		EMIF_INTELLI_PHY_DQS_GATE_OPENING_DELAY_PS * 256 , t_ck);

	return fifo_we_slave_ratio >> 10 | fifo_we_slave_ratio << 1 |
		fifo_we_slave_ratio << 12 | fifo_we_slave_ratio << 23;
}

static u32 get_ext_phy_ctrl_4_intelliphy_4d5(void)
{
	u32 fifo_we_slave_ratio;

	fifo_we_slave_ratio =  DIV_ROUND_CLOSEST(
		EMIF_INTELLI_PHY_DQS_GATE_OPENING_DELAY_PS * 256 , t_ck);

	return fifo_we_slave_ratio >> 9 | fifo_we_slave_ratio << 2 |
		fifo_we_slave_ratio << 13;
}

static u32 get_pwr_mgmt_ctrl(u32 freq, struct emif_data *emif, u32 ip_rev)
{
	u32 pwr_mgmt_ctrl	= 0, timeout;
	u32 lpmode		= EMIF_LP_MODE_SELF_REFRESH;
	u32 timeout_perf	= EMIF_LP_MODE_TIMEOUT_PERFORMANCE;
	u32 timeout_pwr		= EMIF_LP_MODE_TIMEOUT_POWER;
	u32 freq_threshold	= EMIF_LP_MODE_FREQ_THRESHOLD;

	struct emif_custom_configs *cust_cfgs = emif->plat_data->custom_configs;

	if (cust_cfgs && (cust_cfgs->mask & EMIF_CUSTOM_CONFIG_LPMODE)) {
		lpmode		= cust_cfgs->lpmode;
		timeout_perf	= cust_cfgs->lpmode_timeout_performance;
		timeout_pwr	= cust_cfgs->lpmode_timeout_power;
		freq_threshold  = cust_cfgs->lpmode_freq_threshold;
	}

	/* Timeout based on DDR frequency */
	timeout = freq >= freq_threshold ? timeout_perf : timeout_pwr;

	/* The value to be set in register is "log2(timeout) - 3" */
	if (timeout < 16) {
		timeout = 0;
	} else {
		timeout = __fls(timeout) - 3;
		if (timeout & (timeout - 1))
			timeout++;
	}

	switch (lpmode) {
	case EMIF_LP_MODE_CLOCK_STOP:
		pwr_mgmt_ctrl = (timeout << CS_TIM_SHIFT) |
					SR_TIM_MASK | PD_TIM_MASK;
		break;
	case EMIF_LP_MODE_SELF_REFRESH:
		/* Workaround for errata i735 */
		if (timeout < 6)
			timeout = 6;

		pwr_mgmt_ctrl = (timeout << SR_TIM_SHIFT) |
					CS_TIM_MASK | PD_TIM_MASK;
		break;
	case EMIF_LP_MODE_PWR_DN:
		pwr_mgmt_ctrl = (timeout << PD_TIM_SHIFT) |
					CS_TIM_MASK | SR_TIM_MASK;
		break;
	case EMIF_LP_MODE_DISABLE:
	default:
		pwr_mgmt_ctrl = CS_TIM_MASK |
					PD_TIM_MASK | SR_TIM_MASK;
	}

	/* No CS_TIM in EMIF_4D5 */
	if (ip_rev == EMIF_4D5)
		pwr_mgmt_ctrl &= ~CS_TIM_MASK;

	pwr_mgmt_ctrl |= lpmode << LP_MODE_SHIFT;

	return pwr_mgmt_ctrl;
}

/*
 * Program EMIF shadow registers that are not dependent on temperature
 * or voltage
 */
static void setup_registers(struct emif_data *emif, struct emif_regs *regs)
{
	void __iomem	*base = emif->base;

	writel(regs->sdram_tim2_shdw, base + EMIF_SDRAM_TIMING_2_SHDW);
	writel(regs->phy_ctrl_1_shdw, base + EMIF_DDR_PHY_CTRL_1_SHDW);

	/* Settings specific for EMIF4D5 */
	if (emif->plat_data->ip_rev != EMIF_4D5)
		return;
	writel(regs->ext_phy_ctrl_2_shdw, base + EMIF_EXT_PHY_CTRL_2_SHDW);
	writel(regs->ext_phy_ctrl_3_shdw, base + EMIF_EXT_PHY_CTRL_3_SHDW);
	writel(regs->ext_phy_ctrl_4_shdw, base + EMIF_EXT_PHY_CTRL_4_SHDW);
}

/*
 * When voltage ramps dll calibration and forced read idle should
 * happen more often
 */
static void setup_volt_sensitive_regs(struct emif_data *emif,
		struct emif_regs *regs, u32 volt_state)
{
	u32		calib_ctrl;
	void __iomem	*base = emif->base;

	/*
	 * EMIF_READ_IDLE_CTRL in EMIF4D refers to the same register as
	 * EMIF_DLL_CALIB_CTRL in EMIF4D5 and dll_calib_ctrl_shadow_*
	 * is an alias of the respective read_idle_ctrl_shdw_* (members of
	 * a union). So, the below code takes care of both cases
	 */
	if (volt_state == DDR_VOLTAGE_RAMPING)
		calib_ctrl = regs->dll_calib_ctrl_shdw_volt_ramp;
	else
		calib_ctrl = regs->dll_calib_ctrl_shdw_normal;

	writel(calib_ctrl, base + EMIF_DLL_CALIB_CTRL_SHDW);
}

/*
 * setup_temperature_sensitive_regs() - set the timings for temperature
 * sensitive registers. This happens once at initialisation time based
 * on the temperature at boot time and subsequently based on the temperature
 * alert interrupt. Temperature alert can happen when the temperature
 * increases or drops. So this function can have the effect of either
 * derating the timings or going back to nominal values.
 */
static void setup_temperature_sensitive_regs(struct emif_data *emif,
		struct emif_regs *regs)
{
	u32		tim1, tim3, ref_ctrl, type;
	void __iomem	*base = emif->base;
	u32		temperature;

	type = emif->plat_data->device_info->type;

	tim1 = regs->sdram_tim1_shdw;
	tim3 = regs->sdram_tim3_shdw;
	ref_ctrl = regs->ref_ctrl_shdw;

	/* No de-rating for non-lpddr2 devices */
	if (type != DDR_TYPE_LPDDR2_S2 && type != DDR_TYPE_LPDDR2_S4)
		goto out;

	temperature = emif->temperature_level;
	if (temperature == SDRAM_TEMP_HIGH_DERATE_REFRESH) {
		ref_ctrl = regs->ref_ctrl_shdw_derated;
	} else if (temperature == SDRAM_TEMP_HIGH_DERATE_REFRESH_AND_TIMINGS) {
		tim1 = regs->sdram_tim1_shdw_derated;
		tim3 = regs->sdram_tim3_shdw_derated;
		ref_ctrl = regs->ref_ctrl_shdw_derated;
	}

out:
	writel(tim1, base + EMIF_SDRAM_TIMING_1_SHDW);
	writel(tim3, base + EMIF_SDRAM_TIMING_3_SHDW);
	writel(ref_ctrl, base + EMIF_SDRAM_REFRESH_CTRL_SHDW);
}

static void get_default_timings(struct emif_data *emif)
{
	struct emif_platform_data *pd = emif->plat_data;

	pd->timings		= lpddr2_jedec_timings;
	pd->timings_arr_size	= ARRAY_SIZE(lpddr2_jedec_timings);

	dev_warn(emif->dev, "%s: using default timings\n", __func__);
}

static int is_dev_data_valid(u32 type, u32 density, u32 io_width, u32 phy_type,
		u32 ip_rev, struct device *dev)
{
	int valid;

	valid = (type == DDR_TYPE_LPDDR2_S4 ||
			type == DDR_TYPE_LPDDR2_S2)
		&& (density >= DDR_DENSITY_64Mb
			&& density <= DDR_DENSITY_8Gb)
		&& (io_width >= DDR_IO_WIDTH_8
			&& io_width <= DDR_IO_WIDTH_32);

	/* Combinations of EMIF and PHY revisions that we support today */
	switch (ip_rev) {
	case EMIF_4D:
		valid = valid && (phy_type == EMIF_PHY_TYPE_ATTILAPHY);
		break;
	case EMIF_4D5:
		valid = valid && (phy_type == EMIF_PHY_TYPE_INTELLIPHY);
		break;
	default:
		valid = 0;
	}

	if (!valid)
		dev_err(dev, "%s: invalid DDR details\n", __func__);
	return valid;
}

static int is_custom_config_valid(struct emif_custom_configs *cust_cfgs,
		struct device *dev)
{
	int valid = 1;

	if ((cust_cfgs->mask & EMIF_CUSTOM_CONFIG_LPMODE) &&
		(cust_cfgs->lpmode != EMIF_LP_MODE_DISABLE))
		valid = cust_cfgs->lpmode_freq_threshold &&
			cust_cfgs->lpmode_timeout_performance &&
			cust_cfgs->lpmode_timeout_power;

	if (cust_cfgs->mask & EMIF_CUSTOM_CONFIG_TEMP_ALERT_POLL_INTERVAL)
		valid = valid && cust_cfgs->temp_alert_poll_interval_ms;

	if (!valid)
		dev_warn(dev, "%s: invalid custom configs\n", __func__);

	return valid;
}

static struct emif_data *__init_or_module get_device_details(
		struct platform_device *pdev)
{
	u32				size;
	struct emif_data		*emif = NULL;
	struct ddr_device_info		*dev_info;
	struct emif_custom_configs	*cust_cfgs;
	struct emif_platform_data	*pd;
	struct device			*dev;
	void				*temp;

	pd = pdev->dev.platform_data;
	dev = &pdev->dev;

	if (!(pd && pd->device_info && is_dev_data_valid(pd->device_info->type,
			pd->device_info->density, pd->device_info->io_width,
			pd->phy_type, pd->ip_rev, dev))) {
		dev_err(dev, "%s: invalid device data\n", __func__);
		goto error;
	}

	emif	= devm_kzalloc(dev, sizeof(*emif), GFP_KERNEL);
	temp	= devm_kzalloc(dev, sizeof(*pd), GFP_KERNEL);
	dev_info = devm_kzalloc(dev, sizeof(*dev_info), GFP_KERNEL);

	if (!emif || !pd || !dev_info) {
		dev_err(dev, "%s:%d: allocation error\n", __func__, __LINE__);
		goto error;
	}

	memcpy(temp, pd, sizeof(*pd));
	pd = temp;
	memcpy(dev_info, pd->device_info, sizeof(*dev_info));

	pd->device_info		= dev_info;
	emif->plat_data		= pd;
	emif->dev		= dev;
	emif->temperature_level	= SDRAM_TEMP_NOMINAL;

	/*
	 * For EMIF instances other than EMIF1 see if the devices connected
	 * are exactly same as on EMIF1(which is typically the case). If so,
	 * mark it as a duplicate of EMIF1 and skip copying timings data.
	 * This will save some memory and some computation later.
	 */
	emif->duplicate = emif1 && (memcmp(dev_info,
		emif1->plat_data->device_info,
		sizeof(struct ddr_device_info)) == 0);

	if (emif->duplicate) {
		pd->timings = NULL;
		pd->min_tck = NULL;
		goto out;
	} else if (emif1) {
		dev_warn(emif->dev, "%s: Non-symmetric DDR geometry\n",
			__func__);
	}

	/*
	 * Copy custom configs - ignore allocation error, if any, as
	 * custom_configs is not very critical
	 */
	cust_cfgs = pd->custom_configs;
	if (cust_cfgs && is_custom_config_valid(cust_cfgs, dev)) {
		temp = devm_kzalloc(dev, sizeof(*cust_cfgs), GFP_KERNEL);
		if (temp)
			memcpy(temp, cust_cfgs, sizeof(*cust_cfgs));
		else
			dev_warn(dev, "%s:%d: allocation error\n", __func__,
				__LINE__);
		pd->custom_configs = temp;
	}

	/*
	 * Copy timings and min-tck values from platform data. If it is not
	 * available or if memory allocation fails, use JEDEC defaults
	 */
	size = sizeof(struct lpddr2_timings) * pd->timings_arr_size;
	if (pd->timings) {
		temp = devm_kzalloc(dev, size, GFP_KERNEL);
		if (temp) {
			memcpy(temp, pd->timings, sizeof(*pd->timings));
			pd->timings = temp;
		} else {
			dev_warn(dev, "%s:%d: allocation error\n", __func__,
				__LINE__);
			get_default_timings(emif);
		}
	} else {
		get_default_timings(emif);
	}

	if (pd->min_tck) {
		temp = devm_kzalloc(dev, sizeof(*pd->min_tck), GFP_KERNEL);
		if (temp) {
			memcpy(temp, pd->min_tck, sizeof(*pd->min_tck));
			pd->min_tck = temp;
		} else {
			dev_warn(dev, "%s:%d: allocation error\n", __func__,
				__LINE__);
			pd->min_tck = &lpddr2_jedec_min_tck;
		}
	} else {
		pd->min_tck = &lpddr2_jedec_min_tck;
	}

out:
	return emif;

error:
	return NULL;
}

static int __init_or_module emif_probe(struct platform_device *pdev)
{
	struct emif_data	*emif;
	struct resource		*res;

	emif = get_device_details(pdev);
	if (!emif) {
		pr_err("%s: error getting device data\n", __func__);
		goto error;
	}

	list_add(&emif->node, &device_list);
	emif->addressing = get_addressing_table(emif->plat_data->device_info);

	/* Save pointers to each other in emif and device structures */
	emif->dev = &pdev->dev;
	platform_set_drvdata(pdev, emif);

	res = platform_get_resource(pdev, IORESOURCE_MEM, 0);
	if (!res) {
		dev_err(emif->dev, "%s: error getting memory resource\n",
			__func__);
		goto error;
	}

	emif->base = devm_request_and_ioremap(emif->dev, res);
	if (!emif->base) {
		dev_err(emif->dev, "%s: devm_request_and_ioremap() failed\n",
			__func__);
		goto error;
	}

	/* One-time actions taken on probing the first device */
	if (!emif1) {
		emif1 = emif;
		spin_lock_init(&emif_lock);

		/*
		 * TODO: register notifiers for frequency and voltage
		 * change here once the respective frameworks are
		 * available
		 */
	}

	dev_info(&pdev->dev, "%s: device configured with addr = %p\n",
		__func__, emif->base);

	return 0;
error:
	return -ENODEV;
}

static int get_emif_reg_values(struct emif_data *emif, u32 freq,
		struct emif_regs *regs)
{
	u32				cs1_used, ip_rev, phy_type;
	u32				cl, type;
	const struct lpddr2_timings	*timings;
	const struct lpddr2_min_tck	*min_tck;
	const struct ddr_device_info	*device_info;
	const struct lpddr2_addressing	*addressing;
	struct emif_data		*emif_for_calc;
	struct device			*dev;
	const struct emif_custom_configs *custom_configs;

	dev = emif->dev;
	/*
	 * If the devices on this EMIF instance is duplicate of EMIF1,
	 * use EMIF1 details for the calculation
	 */
	emif_for_calc	= emif->duplicate ? emif1 : emif;
	timings		= get_timings_table(emif_for_calc, freq);
	addressing	= emif_for_calc->addressing;
	if (!timings || !addressing) {
		dev_err(dev, "%s: not enough data available for %dHz",
			__func__, freq);
		return -1;
	}

	device_info	= emif_for_calc->plat_data->device_info;
	type		= device_info->type;
	cs1_used	= device_info->cs1_used;
	ip_rev		= emif_for_calc->plat_data->ip_rev;
	phy_type	= emif_for_calc->plat_data->phy_type;

	min_tck		= emif_for_calc->plat_data->min_tck;
	custom_configs	= emif_for_calc->plat_data->custom_configs;

	set_ddr_clk_period(freq);

	regs->ref_ctrl_shdw = get_sdram_ref_ctrl_shdw(freq, addressing);
	regs->sdram_tim1_shdw = get_sdram_tim_1_shdw(timings, min_tck,
			addressing);
	regs->sdram_tim2_shdw = get_sdram_tim_2_shdw(timings, min_tck,
			addressing, type);
	regs->sdram_tim3_shdw = get_sdram_tim_3_shdw(timings, min_tck,
		addressing, type, ip_rev, EMIF_NORMAL_TIMINGS);

	cl = get_cl(emif);

	if (phy_type == EMIF_PHY_TYPE_ATTILAPHY && ip_rev == EMIF_4D) {
		regs->phy_ctrl_1_shdw = get_ddr_phy_ctrl_1_attilaphy_4d(
			timings, freq, cl);
	} else if (phy_type == EMIF_PHY_TYPE_INTELLIPHY && ip_rev == EMIF_4D5) {
		regs->phy_ctrl_1_shdw = get_phy_ctrl_1_intelliphy_4d5(freq, cl);
		regs->ext_phy_ctrl_2_shdw = get_ext_phy_ctrl_2_intelliphy_4d5();
		regs->ext_phy_ctrl_3_shdw = get_ext_phy_ctrl_3_intelliphy_4d5();
		regs->ext_phy_ctrl_4_shdw = get_ext_phy_ctrl_4_intelliphy_4d5();
	} else {
		return -1;
	}

	/* Only timeout values in pwr_mgmt_ctrl_shdw register */
	regs->pwr_mgmt_ctrl_shdw =
		get_pwr_mgmt_ctrl(freq, emif_for_calc, ip_rev) &
		(CS_TIM_MASK | SR_TIM_MASK | PD_TIM_MASK);

	if (ip_rev & EMIF_4D) {
		regs->read_idle_ctrl_shdw_normal =
			get_read_idle_ctrl_shdw(DDR_VOLTAGE_STABLE);

		regs->read_idle_ctrl_shdw_volt_ramp =
			get_read_idle_ctrl_shdw(DDR_VOLTAGE_RAMPING);
	} else if (ip_rev & EMIF_4D5) {
		regs->dll_calib_ctrl_shdw_normal =
			get_dll_calib_ctrl_shdw(DDR_VOLTAGE_STABLE);

		regs->dll_calib_ctrl_shdw_volt_ramp =
			get_dll_calib_ctrl_shdw(DDR_VOLTAGE_RAMPING);
	}

	if (type == DDR_TYPE_LPDDR2_S2 || type == DDR_TYPE_LPDDR2_S4) {
		regs->ref_ctrl_shdw_derated = get_sdram_ref_ctrl_shdw(freq / 4,
			addressing);

		regs->sdram_tim1_shdw_derated =
			get_sdram_tim_1_shdw_derated(timings, min_tck,
				addressing);

		regs->sdram_tim3_shdw_derated = get_sdram_tim_3_shdw(timings,
			min_tck, addressing, type, ip_rev,
			EMIF_DERATED_TIMINGS);
	}

	regs->freq = freq;

	return 0;
}

/*
 * get_regs() - gets the cached emif_regs structure for a given EMIF instance
 * given frequency(freq):
 *
 * As an optimisation, every EMIF instance other than EMIF1 shares the
 * register cache with EMIF1 if the devices connected on this instance
 * are same as that on EMIF1(indicated by the duplicate flag)
 *
 * If we do not have an entry corresponding to the frequency given, we
 * allocate a new entry and calculate the values
 *
 * Upon finding the right reg dump, save it in curr_regs. It can be
 * directly used for thermal de-rating and voltage ramping changes.
 */
static struct emif_regs *get_regs(struct emif_data *emif, u32 freq)
{
	int			i;
	struct emif_regs	**regs_cache;
	struct emif_regs	*regs = NULL;
	struct device		*dev;

	dev = emif->dev;
	if (emif->curr_regs && emif->curr_regs->freq == freq) {
		dev_dbg(dev, "%s: using curr_regs - %u Hz", __func__, freq);
		return emif->curr_regs;
	}

	if (emif->duplicate)
		regs_cache = emif1->regs_cache;
	else
		regs_cache = emif->regs_cache;

	for (i = 0; i < EMIF_MAX_NUM_FREQUENCIES && regs_cache[i]; i++) {
		if (regs_cache[i]->freq == freq) {
			regs = regs_cache[i];
			dev_dbg(dev,
				"%s: reg dump found in reg cache for %u Hz\n",
				__func__, freq);
			break;
		}
	}

	/*
	 * If we don't have an entry for this frequency in the cache create one
	 * and calculate the values
	 */
	if (!regs) {
		regs = devm_kzalloc(emif->dev, sizeof(*regs), GFP_ATOMIC);
		if (!regs)
			return NULL;

		if (get_emif_reg_values(emif, freq, regs)) {
			devm_kfree(emif->dev, regs);
			return NULL;
		}

		/*
		 * Now look for an un-used entry in the cache and save the
		 * newly created struct. If there are no free entries
		 * over-write the last entry
		 */
		for (i = 0; i < EMIF_MAX_NUM_FREQUENCIES && regs_cache[i]; i++)
			;

		if (i >= EMIF_MAX_NUM_FREQUENCIES) {
			dev_warn(dev, "%s: regs_cache full - reusing a slot!!\n",
				__func__);
			i = EMIF_MAX_NUM_FREQUENCIES - 1;
			devm_kfree(emif->dev, regs_cache[i]);
		}
		regs_cache[i] = regs;
	}

	return regs;
}

static void do_volt_notify_handling(struct emif_data *emif, u32 volt_state)
{
	dev_dbg(emif->dev, "%s: voltage notification : %d", __func__,
		volt_state);

	if (!emif->curr_regs) {
		dev_err(emif->dev,
			"%s: volt-notify before registers are ready: %d\n",
			__func__, volt_state);
		return;
	}

	setup_volt_sensitive_regs(emif, emif->curr_regs, volt_state);
}

/*
 * TODO: voltage notify handling should be hooked up to
 * regulator framework as soon as the necessary support
 * is available in mainline kernel. This function is un-used
 * right now.
 */
static void __attribute__((unused)) volt_notify_handling(u32 volt_state)
{
	struct emif_data *emif;

	spin_lock_irqsave(&emif_lock, irq_state);

	list_for_each_entry(emif, &device_list, node)
		do_volt_notify_handling(emif, volt_state);
	do_freq_update();

	spin_unlock_irqrestore(&emif_lock, irq_state);
}

static void do_freq_pre_notify_handling(struct emif_data *emif, u32 new_freq)
{
	struct emif_regs *regs;

	regs = get_regs(emif, new_freq);
	if (!regs)
		return;

	emif->curr_regs = regs;

	/*
	 * Update the shadow registers:
	 * Temperature and voltage-ramp sensitive settings are also configured
	 * in terms of DDR cycles. So, we need to update them too when there
	 * is a freq change
	 */
	dev_dbg(emif->dev, "%s: setting up shadow registers for %uHz",
		__func__, new_freq);
	setup_registers(emif, regs);
	setup_temperature_sensitive_regs(emif, regs);
	setup_volt_sensitive_regs(emif, regs, DDR_VOLTAGE_STABLE);

	/*
	 * Part of workaround for errata i728. See do_freq_update()
	 * for more details
	 */
	if (emif->lpmode == EMIF_LP_MODE_SELF_REFRESH)
		set_lpmode(emif, EMIF_LP_MODE_DISABLE);
}

/*
 * TODO: frequency notify handling should be hooked up to
 * clock framework as soon as the necessary support is
 * available in mainline kernel. This function is un-used
 * right now.
 */
static void __attribute__((unused)) freq_pre_notify_handling(u32 new_freq)
{
	struct emif_data *emif;

	/*
	 * NOTE: we are taking the spin-lock here and releases it
	 * only in post-notifier. This doesn't look good and
	 * Sparse complains about it, but this seems to be
	 * un-avoidable. We need to lock a sequence of events
	 * that is split between EMIF and clock framework.
	 *
	 * 1. EMIF driver updates EMIF timings in shadow registers in the
	 *    frequency pre-notify callback from clock framework
	 * 2. clock framework sets up the registers for the new frequency
	 * 3. clock framework initiates a hw-sequence that updates
	 *    the frequency EMIF timings synchronously.
	 *
	 * All these 3 steps should be performed as an atomic operation
	 * vis-a-vis similar sequence in the EMIF interrupt handler
	 * for temperature events. Otherwise, there could be race
	 * conditions that could result in incorrect EMIF timings for
	 * a given frequency
	 */
	spin_lock_irqsave(&emif_lock, irq_state);

	list_for_each_entry(emif, &device_list, node)
		do_freq_pre_notify_handling(emif, new_freq);
}

static void do_freq_post_notify_handling(struct emif_data *emif)
{
	/*
	 * Part of workaround for errata i728. See do_freq_update()
	 * for more details
	 */
	if (emif->lpmode == EMIF_LP_MODE_SELF_REFRESH)
		set_lpmode(emif, EMIF_LP_MODE_SELF_REFRESH);
}

/*
 * TODO: frequency notify handling should be hooked up to
 * clock framework as soon as the necessary support is
 * available in mainline kernel. This function is un-used
 * right now.
 */
static void __attribute__((unused)) freq_post_notify_handling(void)
{
	struct emif_data *emif;

	list_for_each_entry(emif, &device_list, node)
		do_freq_post_notify_handling(emif);

	/*
	 * Lock is done in pre-notify handler. See freq_pre_notify_handling()
	 * for more details
	 */
	spin_unlock_irqrestore(&emif_lock, irq_state);
}

static struct platform_driver emif_driver = {
	.driver = {
		.name = "emif",
	},
};

static int __init_or_module emif_register(void)
{
	return platform_driver_probe(&emif_driver, emif_probe);
}

static void __exit emif_unregister(void)
{
	platform_driver_unregister(&emif_driver);
}

module_init(emif_register);
module_exit(emif_unregister);
MODULE_DESCRIPTION("TI EMIF SDRAM Controller Driver");
MODULE_LICENSE("GPL");
MODULE_ALIAS("platform:emif");
MODULE_AUTHOR("Texas Instruments Inc");