sequencer.c

/*
 * Copyright Altera Corporation (C) 2012-2015
 *
 * SPDX-License-Identifier:    BSD-3-Clause
 */

#include <common.h>
#include <asm/io.h>
#include <asm/arch/sdram.h>
#include "sequencer.h"
#include "sequencer_auto.h"
#include "sequencer_auto_ac_init.h"
#include "sequencer_auto_inst_init.h"
#include "sequencer_defines.h"

static struct socfpga_sdr_rw_load_manager *sdr_rw_load_mgr_regs =
	(struct socfpga_sdr_rw_load_manager *)(SDR_PHYGRP_RWMGRGRP_ADDRESS | 0x800);

static struct socfpga_sdr_rw_load_jump_manager *sdr_rw_load_jump_mgr_regs =
	(struct socfpga_sdr_rw_load_jump_manager *)(SDR_PHYGRP_RWMGRGRP_ADDRESS | 0xC00);

static struct socfpga_sdr_reg_file *sdr_reg_file =
	(struct socfpga_sdr_reg_file *)SDR_PHYGRP_REGFILEGRP_ADDRESS;

static struct socfpga_sdr_scc_mgr *sdr_scc_mgr =
	(struct socfpga_sdr_scc_mgr *)(SDR_PHYGRP_SCCGRP_ADDRESS | 0xe00);

static struct socfpga_phy_mgr_cmd *phy_mgr_cmd =
	(struct socfpga_phy_mgr_cmd *)SDR_PHYGRP_PHYMGRGRP_ADDRESS;

static struct socfpga_phy_mgr_cfg *phy_mgr_cfg =
	(struct socfpga_phy_mgr_cfg *)(SDR_PHYGRP_PHYMGRGRP_ADDRESS | 0x40);

static struct socfpga_data_mgr *data_mgr =
	(struct socfpga_data_mgr *)SDR_PHYGRP_DATAMGRGRP_ADDRESS;

static struct socfpga_sdr_ctrl *sdr_ctrl =
	(struct socfpga_sdr_ctrl *)SDR_CTRLGRP_ADDRESS;

#define DELTA_D		1

/*
 * In order to reduce ROM size, most of the selectable calibration steps are
 * decided at compile time based on the user's calibration mode selection,
 * as captured by the STATIC_CALIB_STEPS selection below.
 *
 * However, to support simulation-time selection of fast simulation mode, where
 * we skip everything except the bare minimum, we need a few of the steps to
 * be dynamic.  In those cases, we either use the DYNAMIC_CALIB_STEPS for the
 * check, which is based on the rtl-supplied value, or we dynamically compute
 * the value to use based on the dynamically-chosen calibration mode
 */

#define DLEVEL 0
#define STATIC_IN_RTL_SIM 0
#define STATIC_SKIP_DELAY_LOOPS 0

#define STATIC_CALIB_STEPS (STATIC_IN_RTL_SIM | CALIB_SKIP_FULL_TEST | \
	STATIC_SKIP_DELAY_LOOPS)

/* calibration steps requested by the rtl */
uint16_t dyn_calib_steps;

/*
 * To make CALIB_SKIP_DELAY_LOOPS a dynamic conditional option
 * instead of static, we use boolean logic to select between
 * non-skip and skip values
 *
 * The mask is set to include all bits when not-skipping, but is
 * zero when skipping
 */

uint16_t skip_delay_mask;	/* mask off bits when skipping/not-skipping */

#define SKIP_DELAY_LOOP_VALUE_OR_ZERO(non_skip_value) \
	((non_skip_value) & skip_delay_mask)

struct gbl_type *gbl;
struct param_type *param;
uint32_t curr_shadow_reg;

static uint32_t rw_mgr_mem_calibrate_write_test(uint32_t rank_bgn,
	uint32_t write_group, uint32_t use_dm,
	uint32_t all_correct, uint32_t *bit_chk, uint32_t all_ranks);

static void set_failing_group_stage(uint32_t group, uint32_t stage,
	uint32_t substage)
{
	/*
	 * Only set the global stage if there was not been any other
	 * failing group
	 */
	if (gbl->error_stage == CAL_STAGE_NIL)	{
		gbl->error_substage = substage;
		gbl->error_stage = stage;
		gbl->error_group = group;
	}
}

static void reg_file_set_group(u16 set_group)
{
	clrsetbits_le32(&sdr_reg_file->cur_stage, 0xffff0000, set_group << 16);
}

static void reg_file_set_stage(u8 set_stage)
{
	clrsetbits_le32(&sdr_reg_file->cur_stage, 0xffff, set_stage & 0xff);
}

static void reg_file_set_sub_stage(u8 set_sub_stage)
{
	set_sub_stage &= 0xff;
	clrsetbits_le32(&sdr_reg_file->cur_stage, 0xff00, set_sub_stage << 8);
}

static void initialize(void)
{
	debug("%s:%d\n", __func__, __LINE__);
	/* USER calibration has control over path to memory */
	/*
	 * In Hard PHY this is a 2-bit control:
	 * 0: AFI Mux Select
	 * 1: DDIO Mux Select
	 */
	writel(0x3, &phy_mgr_cfg->mux_sel);

	/* USER memory clock is not stable we begin initialization  */
	writel(0, &phy_mgr_cfg->reset_mem_stbl);

	/* USER calibration status all set to zero */
	writel(0, &phy_mgr_cfg->cal_status);

	writel(0, &phy_mgr_cfg->cal_debug_info);

	if ((dyn_calib_steps & CALIB_SKIP_ALL) != CALIB_SKIP_ALL) {
		param->read_correct_mask_vg  = ((uint32_t)1 <<
			(RW_MGR_MEM_DQ_PER_READ_DQS /
			RW_MGR_MEM_VIRTUAL_GROUPS_PER_READ_DQS)) - 1;
		param->write_correct_mask_vg = ((uint32_t)1 <<
			(RW_MGR_MEM_DQ_PER_READ_DQS /
			RW_MGR_MEM_VIRTUAL_GROUPS_PER_READ_DQS)) - 1;
		param->read_correct_mask     = ((uint32_t)1 <<
			RW_MGR_MEM_DQ_PER_READ_DQS) - 1;
		param->write_correct_mask    = ((uint32_t)1 <<
			RW_MGR_MEM_DQ_PER_WRITE_DQS) - 1;
		param->dm_correct_mask       = ((uint32_t)1 <<
			(RW_MGR_MEM_DATA_WIDTH / RW_MGR_MEM_DATA_MASK_WIDTH))
			- 1;
	}
}

static void set_rank_and_odt_mask(uint32_t rank, uint32_t odt_mode)
{
	uint32_t odt_mask_0 = 0;
	uint32_t odt_mask_1 = 0;
	uint32_t cs_and_odt_mask;

	if (odt_mode == RW_MGR_ODT_MODE_READ_WRITE) {
		if (RW_MGR_MEM_NUMBER_OF_RANKS == 1) {
			/*
			 * 1 Rank
			 * Read: ODT = 0
			 * Write: ODT = 1
			 */
			odt_mask_0 = 0x0;
			odt_mask_1 = 0x1;
		} else if (RW_MGR_MEM_NUMBER_OF_RANKS == 2) {
			/* 2 Ranks */
			if (RW_MGR_MEM_NUMBER_OF_CS_PER_DIMM == 1) {
				/* - Dual-Slot , Single-Rank
				 * (1 chip-select per DIMM)
				 * OR
				 * - RDIMM, 4 total CS (2 CS per DIMM)
				 * means 2 DIMM
				 * Since MEM_NUMBER_OF_RANKS is 2 they are
				 * both single rank
				 * with 2 CS each (special for RDIMM)
				 * Read: Turn on ODT on the opposite rank
				 * Write: Turn on ODT on all ranks
				 */
				odt_mask_0 = 0x3 & ~(1 << rank);
				odt_mask_1 = 0x3;
			} else {
				/*
				 * USER - Single-Slot , Dual-rank DIMMs
				 * (2 chip-selects per DIMM)
				 * USER Read: Turn on ODT off on all ranks
				 * USER Write: Turn on ODT on active rank
				 */
				odt_mask_0 = 0x0;
				odt_mask_1 = 0x3 & (1 << rank);
			}
		} else {
			/* 4 Ranks
			 * Read:
			 * ----------+-----------------------+
			 *           |                       |
			 *           |         ODT           |
			 * Read From +-----------------------+
			 *   Rank    |  3  |  2  |  1  |  0  |
			 * ----------+-----+-----+-----+-----+
			 *     0     |  0  |  1  |  0  |  0  |
			 *     1     |  1  |  0  |  0  |  0  |
			 *     2     |  0  |  0  |  0  |  1  |
			 *     3     |  0  |  0  |  1  |  0  |
			 * ----------+-----+-----+-----+-----+
			 *
			 * Write:
			 * ----------+-----------------------+
			 *           |                       |
			 *           |         ODT           |
			 * Write To  +-----------------------+
			 *   Rank    |  3  |  2  |  1  |  0  |
			 * ----------+-----+-----+-----+-----+
			 *     0     |  0  |  1  |  0  |  1  |
			 *     1     |  1  |  0  |  1  |  0  |
			 *     2     |  0  |  1  |  0  |  1  |
			 *     3     |  1  |  0  |  1  |  0  |
			 * ----------+-----+-----+-----+-----+
			 */
			switch (rank) {
			case 0:
				odt_mask_0 = 0x4;
				odt_mask_1 = 0x5;
				break;
			case 1:
				odt_mask_0 = 0x8;
				odt_mask_1 = 0xA;
				break;
			case 2:
				odt_mask_0 = 0x1;
				odt_mask_1 = 0x5;
				break;
			case 3:
				odt_mask_0 = 0x2;
				odt_mask_1 = 0xA;
				break;
			}
		}
	} else {
		odt_mask_0 = 0x0;
		odt_mask_1 = 0x0;
	}

	cs_and_odt_mask =
		(0xFF & ~(1 << rank)) |
		((0xFF & odt_mask_0) << 8) |
		((0xFF & odt_mask_1) << 16);
	writel(cs_and_odt_mask, SDR_PHYGRP_RWMGRGRP_ADDRESS |
				RW_MGR_SET_CS_AND_ODT_MASK_OFFSET);
}

/**
 * scc_mgr_set() - Set SCC Manager register
 * @off:	Base offset in SCC Manager space
 * @grp:	Read/Write group
 * @val:	Value to be set
 *
 * This function sets the SCC Manager (Scan Chain Control Manager) register.
 */
static void scc_mgr_set(u32 off, u32 grp, u32 val)
{
	writel(val, SDR_PHYGRP_SCCGRP_ADDRESS | off | (grp << 2));
}

/**
 * scc_mgr_initialize() - Initialize SCC Manager registers
 *
 * Initialize SCC Manager registers.
 */
static void scc_mgr_initialize(void)
{
	/*
	 * Clear register file for HPS. 16 (2^4) is the size of the
	 * full register file in the scc mgr:
	 *	RFILE_DEPTH = 1 + log2(MEM_DQ_PER_DQS + 1 + MEM_DM_PER_DQS +
	 *                             MEM_IF_READ_DQS_WIDTH - 1);
	 */
	int i;

	for (i = 0; i < 16; i++) {
		debug_cond(DLEVEL == 1, "%s:%d: Clearing SCC RFILE index %u\n",
			   __func__, __LINE__, i);
		scc_mgr_set(SCC_MGR_HHP_RFILE_OFFSET, 0, i);
	}
}

static void scc_mgr_set_dqdqs_output_phase(uint32_t write_group, uint32_t phase)
{
	scc_mgr_set(SCC_MGR_DQDQS_OUT_PHASE_OFFSET, write_group, phase);
}

static void scc_mgr_set_dqs_bus_in_delay(uint32_t read_group, uint32_t delay)
{
	scc_mgr_set(SCC_MGR_DQS_IN_DELAY_OFFSET, read_group, delay);
}

static void scc_mgr_set_dqs_en_phase(uint32_t read_group, uint32_t phase)
{
	scc_mgr_set(SCC_MGR_DQS_EN_PHASE_OFFSET, read_group, phase);
}

static void scc_mgr_set_dqs_en_delay(uint32_t read_group, uint32_t delay)
{
	scc_mgr_set(SCC_MGR_DQS_EN_DELAY_OFFSET, read_group, delay);
}

static void scc_mgr_set_dqs_io_in_delay(uint32_t delay)
{
	scc_mgr_set(SCC_MGR_IO_IN_DELAY_OFFSET, RW_MGR_MEM_DQ_PER_WRITE_DQS,
		    delay);
}

static void scc_mgr_set_dq_in_delay(uint32_t dq_in_group, uint32_t delay)
{
	scc_mgr_set(SCC_MGR_IO_IN_DELAY_OFFSET, dq_in_group, delay);
}

static void scc_mgr_set_dq_out1_delay(uint32_t dq_in_group, uint32_t delay)
{
	scc_mgr_set(SCC_MGR_IO_OUT1_DELAY_OFFSET, dq_in_group, delay);
}

static void scc_mgr_set_dqs_out1_delay(uint32_t delay)
{
	scc_mgr_set(SCC_MGR_IO_OUT1_DELAY_OFFSET, RW_MGR_MEM_DQ_PER_WRITE_DQS,
		    delay);
}

static void scc_mgr_set_dm_out1_delay(uint32_t dm, uint32_t delay)
{
	scc_mgr_set(SCC_MGR_IO_OUT1_DELAY_OFFSET,
		    RW_MGR_MEM_DQ_PER_WRITE_DQS + 1 + dm,
		    delay);
}

/* load up dqs config settings */
static void scc_mgr_load_dqs(uint32_t dqs)
{
	writel(dqs, &sdr_scc_mgr->dqs_ena);
}

/* load up dqs io config settings */
static void scc_mgr_load_dqs_io(void)
{
	writel(0, &sdr_scc_mgr->dqs_io_ena);
}

/* load up dq config settings */
static void scc_mgr_load_dq(uint32_t dq_in_group)
{
	writel(dq_in_group, &sdr_scc_mgr->dq_ena);
}

/* load up dm config settings */
static void scc_mgr_load_dm(uint32_t dm)
{
	writel(dm, &sdr_scc_mgr->dm_ena);
}

/**
 * scc_mgr_set_all_ranks() - Set SCC Manager register for all ranks
 * @off:	Base offset in SCC Manager space
 * @grp:	Read/Write group
 * @val:	Value to be set
 * @update:	If non-zero, trigger SCC Manager update for all ranks
 *
 * This function sets the SCC Manager (Scan Chain Control Manager) register
 * and optionally triggers the SCC update for all ranks.
 */
static void scc_mgr_set_all_ranks(const u32 off, const u32 grp, const u32 val,
				  const int update)
{
	u32 r;

	for (r = 0; r < RW_MGR_MEM_NUMBER_OF_RANKS;
	     r += NUM_RANKS_PER_SHADOW_REG) {
		scc_mgr_set(off, grp, val);

		if (update || (r == 0)) {
			writel(grp, &sdr_scc_mgr->dqs_ena);
			writel(0, &sdr_scc_mgr->update);
		}
	}
}

static void scc_mgr_set_dqs_en_phase_all_ranks(u32 read_group, u32 phase)
{
	/*
	 * USER although the h/w doesn't support different phases per
	 * shadow register, for simplicity our scc manager modeling
	 * keeps different phase settings per shadow reg, and it's
	 * important for us to keep them in sync to match h/w.
	 * for efficiency, the scan chain update should occur only
	 * once to sr0.
	 */
	scc_mgr_set_all_ranks(SCC_MGR_DQS_EN_PHASE_OFFSET,
			      read_group, phase, 0);
}

static void scc_mgr_set_dqdqs_output_phase_all_ranks(uint32_t write_group,
						     uint32_t phase)
{
	/*
	 * USER although the h/w doesn't support different phases per
	 * shadow register, for simplicity our scc manager modeling
	 * keeps different phase settings per shadow reg, and it's
	 * important for us to keep them in sync to match h/w.
	 * for efficiency, the scan chain update should occur only
	 * once to sr0.
	 */
	scc_mgr_set_all_ranks(SCC_MGR_DQDQS_OUT_PHASE_OFFSET,
			      write_group, phase, 0);
}

static void scc_mgr_set_dqs_en_delay_all_ranks(uint32_t read_group,
					       uint32_t delay)
{
	/*
	 * In shadow register mode, the T11 settings are stored in
	 * registers in the core, which are updated by the DQS_ENA
	 * signals. Not issuing the SCC_MGR_UPD command allows us to
	 * save lots of rank switching overhead, by calling
	 * select_shadow_regs_for_update with update_scan_chains
	 * set to 0.
	 */
	scc_mgr_set_all_ranks(SCC_MGR_DQS_EN_DELAY_OFFSET,
			      read_group, delay, 1);
	writel(0, &sdr_scc_mgr->update);
}

/**
 * scc_mgr_set_oct_out1_delay() - Set OCT output delay
 * @write_group:	Write group
 * @delay:		Delay value
 *
 * This function sets the OCT output delay in SCC manager.
 */
static void scc_mgr_set_oct_out1_delay(const u32 write_group, const u32 delay)
{
	const int ratio = RW_MGR_MEM_IF_READ_DQS_WIDTH /
			  RW_MGR_MEM_IF_WRITE_DQS_WIDTH;
	const int base = write_group * ratio;
	int i;
	/*
	 * Load the setting in the SCC manager
	 * Although OCT affects only write data, the OCT delay is controlled
	 * by the DQS logic block which is instantiated once per read group.
	 * For protocols where a write group consists of multiple read groups,
	 * the setting must be set multiple times.
	 */
	for (i = 0; i < ratio; i++)
		scc_mgr_set(SCC_MGR_OCT_OUT1_DELAY_OFFSET, base + i, delay);
}

/**
 * scc_mgr_set_hhp_extras() - Set HHP extras.
 *
 * Load the fixed setting in the SCC manager HHP extras.
 */
static void scc_mgr_set_hhp_extras(void)
{
	/*
	 * Load the fixed setting in the SCC manager
	 * bits: 0:0 = 1'b1	- DQS bypass
	 * bits: 1:1 = 1'b1	- DQ bypass
	 * bits: 4:2 = 3'b001	- rfifo_mode
	 * bits: 6:5 = 2'b01	- rfifo clock_select
	 * bits: 7:7 = 1'b0	- separate gating from ungating setting
	 * bits: 8:8 = 1'b0	- separate OE from Output delay setting
	 */
	const u32 value = (0 << 8) | (0 << 7) | (1 << 5) |
			  (1 << 2) | (1 << 1) | (1 << 0);
	const u32 addr = SDR_PHYGRP_SCCGRP_ADDRESS |
			 SCC_MGR_HHP_GLOBALS_OFFSET |
			 SCC_MGR_HHP_EXTRAS_OFFSET;

	debug_cond(DLEVEL == 1, "%s:%d Setting HHP Extras\n",
		   __func__, __LINE__);
	writel(value, addr);
	debug_cond(DLEVEL == 1, "%s:%d Done Setting HHP Extras\n",
		   __func__, __LINE__);
}

/**
 * scc_mgr_zero_all() - Zero all DQS config
 *
 * Zero all DQS config.
 */
static void scc_mgr_zero_all(void)
{
	int i, r;

	/*
	 * USER Zero all DQS config settings, across all groups and all
	 * shadow registers
	 */
	for (r = 0; r < RW_MGR_MEM_NUMBER_OF_RANKS;
	     r += NUM_RANKS_PER_SHADOW_REG) {
		for (i = 0; i < RW_MGR_MEM_IF_READ_DQS_WIDTH; i++) {
			/*
			 * The phases actually don't exist on a per-rank basis,
			 * but there's no harm updating them several times, so
			 * let's keep the code simple.
			 */
			scc_mgr_set_dqs_bus_in_delay(i, IO_DQS_IN_RESERVE);
			scc_mgr_set_dqs_en_phase(i, 0);
			scc_mgr_set_dqs_en_delay(i, 0);
		}

		for (i = 0; i < RW_MGR_MEM_IF_WRITE_DQS_WIDTH; i++) {
			scc_mgr_set_dqdqs_output_phase(i, 0);
			/* Arria V/Cyclone V don't have out2. */
			scc_mgr_set_oct_out1_delay(i, IO_DQS_OUT_RESERVE);
		}
	}

	/* Multicast to all DQS group enables. */
	writel(0xff, &sdr_scc_mgr->dqs_ena);
	writel(0, &sdr_scc_mgr->update);
}

/**
 * scc_set_bypass_mode() - Set bypass mode and trigger SCC update
 * @write_group:	Write group
 *
 * Set bypass mode and trigger SCC update.
 */
static void scc_set_bypass_mode(const u32 write_group)
{
	/* Multicast to all DQ enables. */
	writel(0xff, &sdr_scc_mgr->dq_ena);
	writel(0xff, &sdr_scc_mgr->dm_ena);

	/* Update current DQS IO enable. */
	writel(0, &sdr_scc_mgr->dqs_io_ena);

	/* Update the DQS logic. */
	writel(write_group, &sdr_scc_mgr->dqs_ena);

	/* Hit update. */
	writel(0, &sdr_scc_mgr->update);
}

/**
 * scc_mgr_load_dqs_for_write_group() - Load DQS settings for Write Group
 * @write_group:	Write group
 *
 * Load DQS settings for Write Group, do not trigger SCC update.
 */
static void scc_mgr_load_dqs_for_write_group(const u32 write_group)
{
	const int ratio = RW_MGR_MEM_IF_READ_DQS_WIDTH /
			  RW_MGR_MEM_IF_WRITE_DQS_WIDTH;
	const int base = write_group * ratio;
	int i;
	/*
	 * Load the setting in the SCC manager
	 * Although OCT affects only write data, the OCT delay is controlled
	 * by the DQS logic block which is instantiated once per read group.
	 * For protocols where a write group consists of multiple read groups,
	 * the setting must be set multiple times.
	 */
	for (i = 0; i < ratio; i++)
		writel(base + i, &sdr_scc_mgr->dqs_ena);
}

/**
 * scc_mgr_zero_group() - Zero all configs for a group
 *
 * Zero DQ, DM, DQS and OCT configs for a group.
 */
static void scc_mgr_zero_group(const u32 write_group, const int out_only)
{
	int i, r;

	for (r = 0; r < RW_MGR_MEM_NUMBER_OF_RANKS;
	     r += NUM_RANKS_PER_SHADOW_REG) {
		/* Zero all DQ config settings. */
		for (i = 0; i < RW_MGR_MEM_DQ_PER_WRITE_DQS; i++) {
			scc_mgr_set_dq_out1_delay(i, 0);
			if (!out_only)
				scc_mgr_set_dq_in_delay(i, 0);
		}

		/* Multicast to all DQ enables. */
		writel(0xff, &sdr_scc_mgr->dq_ena);

		/* Zero all DM config settings. */
		for (i = 0; i < RW_MGR_NUM_DM_PER_WRITE_GROUP; i++)
			scc_mgr_set_dm_out1_delay(i, 0);

		/* Multicast to all DM enables. */
		writel(0xff, &sdr_scc_mgr->dm_ena);

		/* Zero all DQS IO settings. */
		if (!out_only)
			scc_mgr_set_dqs_io_in_delay(0);

		/* Arria V/Cyclone V don't have out2. */
		scc_mgr_set_dqs_out1_delay(IO_DQS_OUT_RESERVE);
		scc_mgr_set_oct_out1_delay(write_group, IO_DQS_OUT_RESERVE);
		scc_mgr_load_dqs_for_write_group(write_group);

		/* Multicast to all DQS IO enables (only 1 in total). */
		writel(0, &sdr_scc_mgr->dqs_io_ena);

		/* Hit update to zero everything. */
		writel(0, &sdr_scc_mgr->update);
	}
}

/*
 * apply and load a particular input delay for the DQ pins in a group
 * group_bgn is the index of the first dq pin (in the write group)
 */
static void scc_mgr_apply_group_dq_in_delay(uint32_t group_bgn, uint32_t delay)
{
	uint32_t i, p;

	for (i = 0, p = group_bgn; i < RW_MGR_MEM_DQ_PER_READ_DQS; i++, p++) {
		scc_mgr_set_dq_in_delay(p, delay);
		scc_mgr_load_dq(p);
	}
}

/**
 * scc_mgr_apply_group_dq_out1_delay() - Apply and load an output delay for the DQ pins in a group
 * @delay:		Delay value
 *
 * Apply and load a particular output delay for the DQ pins in a group.
 */
static void scc_mgr_apply_group_dq_out1_delay(const u32 delay)
{
	int i;

	for (i = 0; i < RW_MGR_MEM_DQ_PER_WRITE_DQS; i++) {
		scc_mgr_set_dq_out1_delay(i, delay);
		scc_mgr_load_dq(i);
	}
}

/* apply and load a particular output delay for the DM pins in a group */
static void scc_mgr_apply_group_dm_out1_delay(uint32_t delay1)
{
	uint32_t i;

	for (i = 0; i < RW_MGR_NUM_DM_PER_WRITE_GROUP; i++) {
		scc_mgr_set_dm_out1_delay(i, delay1);
		scc_mgr_load_dm(i);
	}
}


/* apply and load delay on both DQS and OCT out1 */
static void scc_mgr_apply_group_dqs_io_and_oct_out1(uint32_t write_group,
						    uint32_t delay)
{
	scc_mgr_set_dqs_out1_delay(delay);
	scc_mgr_load_dqs_io();

	scc_mgr_set_oct_out1_delay(write_group, delay);
	scc_mgr_load_dqs_for_write_group(write_group);
}

/**
 * scc_mgr_apply_group_all_out_delay_add() - Apply a delay to the entire output side: DQ, DM, DQS, OCT
 * @write_group:	Write group
 * @delay:		Delay value
 *
 * Apply a delay to the entire output side: DQ, DM, DQS, OCT.
 */
static void scc_mgr_apply_group_all_out_delay_add(const u32 write_group,
						  const u32 delay)
{
	u32 i, new_delay;

	/* DQ shift */
	for (i = 0; i < RW_MGR_MEM_DQ_PER_WRITE_DQS; i++)
		scc_mgr_load_dq(i);

	/* DM shift */
	for (i = 0; i < RW_MGR_NUM_DM_PER_WRITE_GROUP; i++)
		scc_mgr_load_dm(i);

	/* DQS shift */
	new_delay = READ_SCC_DQS_IO_OUT2_DELAY + delay;
	if (new_delay > IO_IO_OUT2_DELAY_MAX) {
		debug_cond(DLEVEL == 1,
			   "%s:%d (%u, %u) DQS: %u > %d; adding %u to OUT1\n",
			   __func__, __LINE__, write_group, delay, new_delay,
			   IO_IO_OUT2_DELAY_MAX,
			   new_delay - IO_IO_OUT2_DELAY_MAX);
		new_delay -= IO_IO_OUT2_DELAY_MAX;
		scc_mgr_set_dqs_out1_delay(new_delay);
	}

	scc_mgr_load_dqs_io();

	/* OCT shift */
	new_delay = READ_SCC_OCT_OUT2_DELAY + delay;
	if (new_delay > IO_IO_OUT2_DELAY_MAX) {
		debug_cond(DLEVEL == 1,
			   "%s:%d (%u, %u) DQS: %u > %d; adding %u to OUT1\n",
			   __func__, __LINE__, write_group, delay,
			   new_delay, IO_IO_OUT2_DELAY_MAX,
			   new_delay - IO_IO_OUT2_DELAY_MAX);
		new_delay -= IO_IO_OUT2_DELAY_MAX;
		scc_mgr_set_oct_out1_delay(write_group, new_delay);
	}

	scc_mgr_load_dqs_for_write_group(write_group);
}

/**
 * scc_mgr_apply_group_all_out_delay_add() - Apply a delay to the entire output side to all ranks
 * @write_group:	Write group
 * @delay:		Delay value
 *
 * Apply a delay to the entire output side (DQ, DM, DQS, OCT) to all ranks.
 */
static void
scc_mgr_apply_group_all_out_delay_add_all_ranks(const u32 write_group,
						const u32 delay)
{
	int r;

	for (r = 0; r < RW_MGR_MEM_NUMBER_OF_RANKS;
	     r += NUM_RANKS_PER_SHADOW_REG) {
		scc_mgr_apply_group_all_out_delay_add(write_group, delay);
		writel(0, &sdr_scc_mgr->update);
	}
}

/* optimization used to recover some slots in ddr3 inst_rom */
/* could be applied to other protocols if we wanted to */
static void set_jump_as_return(void)
{
	/*
	 * to save space, we replace return with jump to special shared
	 * RETURN instruction so we set the counter to large value so that
	 * we always jump
	 */
	writel(0xff, &sdr_rw_load_mgr_regs->load_cntr0);
	writel(RW_MGR_RETURN, &sdr_rw_load_jump_mgr_regs->load_jump_add0);
}

/*
 * should always use constants as argument to ensure all computations are
 * performed at compile time
 */
static void delay_for_n_mem_clocks(const uint32_t clocks)
{
	uint32_t afi_clocks;
	uint8_t inner = 0;
	uint8_t outer = 0;
	uint16_t c_loop = 0;

	debug("%s:%d: clocks=%u ... start\n", __func__, __LINE__, clocks);


	afi_clocks = (clocks + AFI_RATE_RATIO-1) / AFI_RATE_RATIO;
	/* scale (rounding up) to get afi clocks */

	/*
	 * Note, we don't bother accounting for being off a little bit
	 * because of a few extra instructions in outer loops
	 * Note, the loops have a test at the end, and do the test before
	 * the decrement, and so always perform the loop
	 * 1 time more than the counter value
	 */
	if (afi_clocks == 0) {
		;
	} else if (afi_clocks <= 0x100) {
		inner = afi_clocks-1;
		outer = 0;
		c_loop = 0;
	} else if (afi_clocks <= 0x10000) {
		inner = 0xff;
		outer = (afi_clocks-1) >> 8;
		c_loop = 0;
	} else {
		inner = 0xff;
		outer = 0xff;
		c_loop = (afi_clocks-1) >> 16;
	}

	/*
	 * rom instructions are structured as follows:
	 *
	 *    IDLE_LOOP2: jnz cntr0, TARGET_A
	 *    IDLE_LOOP1: jnz cntr1, TARGET_B
	 *                return
	 *
	 * so, when doing nested loops, TARGET_A is set to IDLE_LOOP2, and
	 * TARGET_B is set to IDLE_LOOP2 as well
	 *
	 * if we have no outer loop, though, then we can use IDLE_LOOP1 only,
	 * and set TARGET_B to IDLE_LOOP1 and we skip IDLE_LOOP2 entirely
	 *
	 * a little confusing, but it helps save precious space in the inst_rom
	 * and sequencer rom and keeps the delays more accurate and reduces
	 * overhead
	 */
	if (afi_clocks <= 0x100) {
		writel(SKIP_DELAY_LOOP_VALUE_OR_ZERO(inner),
			&sdr_rw_load_mgr_regs->load_cntr1);

		writel(RW_MGR_IDLE_LOOP1,
			&sdr_rw_load_jump_mgr_regs->load_jump_add1);

		writel(RW_MGR_IDLE_LOOP1, SDR_PHYGRP_RWMGRGRP_ADDRESS |
					  RW_MGR_RUN_SINGLE_GROUP_OFFSET);
	} else {
		writel(SKIP_DELAY_LOOP_VALUE_OR_ZERO(inner),
			&sdr_rw_load_mgr_regs->load_cntr0);

		writel(SKIP_DELAY_LOOP_VALUE_OR_ZERO(outer),
			&sdr_rw_load_mgr_regs->load_cntr1);

		writel(RW_MGR_IDLE_LOOP2,
			&sdr_rw_load_jump_mgr_regs->load_jump_add0);

		writel(RW_MGR_IDLE_LOOP2,
			&sdr_rw_load_jump_mgr_regs->load_jump_add1);

		/* hack to get around compiler not being smart enough */
		if (afi_clocks <= 0x10000) {
			/* only need to run once */
			writel(RW_MGR_IDLE_LOOP2, SDR_PHYGRP_RWMGRGRP_ADDRESS |
						  RW_MGR_RUN_SINGLE_GROUP_OFFSET);
		} else {
			do {
				writel(RW_MGR_IDLE_LOOP2,
					SDR_PHYGRP_RWMGRGRP_ADDRESS |
					RW_MGR_RUN_SINGLE_GROUP_OFFSET);
			} while (c_loop-- != 0);
		}
	}
	debug("%s:%d clocks=%u ... end\n", __func__, __LINE__, clocks);
}

static void rw_mgr_mem_initialize(void)
{
	uint32_t r;
	uint32_t grpaddr = SDR_PHYGRP_RWMGRGRP_ADDRESS |
			   RW_MGR_RUN_SINGLE_GROUP_OFFSET;

	debug("%s:%d\n", __func__, __LINE__);

	/* The reset / cke part of initialization is broadcasted to all ranks */
	writel(RW_MGR_RANK_ALL, SDR_PHYGRP_RWMGRGRP_ADDRESS |
				RW_MGR_SET_CS_AND_ODT_MASK_OFFSET);

	/*
	 * Here's how you load register for a loop
	 * Counters are located @ 0x800
	 * Jump address are located @ 0xC00
	 * For both, registers 0 to 3 are selected using bits 3 and 2, like
	 * in 0x800, 0x804, 0x808, 0x80C and 0xC00, 0xC04, 0xC08, 0xC0C
	 * I know this ain't pretty, but Avalon bus throws away the 2 least
	 * significant bits
	 */

	/* start with memory RESET activated */

	/* tINIT = 200us */

	/*
	 * 200us @ 266MHz (3.75 ns) ~ 54000 clock cycles
	 * If a and b are the number of iteration in 2 nested loops
	 * it takes the following number of cycles to complete the operation:
	 * number_of_cycles = ((2 + n) * a + 2) * b
	 * where n is the number of instruction in the inner loop
	 * One possible solution is n = 0 , a = 256 , b = 106 => a = FF,
	 * b = 6A
	 */

	/* Load counters */
	writel(SKIP_DELAY_LOOP_VALUE_OR_ZERO(SEQ_TINIT_CNTR0_VAL),
	       &sdr_rw_load_mgr_regs->load_cntr0);
	writel(SKIP_DELAY_LOOP_VALUE_OR_ZERO(SEQ_TINIT_CNTR1_VAL),
	       &sdr_rw_load_mgr_regs->load_cntr1);
	writel(SKIP_DELAY_LOOP_VALUE_OR_ZERO(SEQ_TINIT_CNTR2_VAL),
	       &sdr_rw_load_mgr_regs->load_cntr2);

	/* Load jump address */
	writel(RW_MGR_INIT_RESET_0_CKE_0,
		&sdr_rw_load_jump_mgr_regs->load_jump_add0);
	writel(RW_MGR_INIT_RESET_0_CKE_0,
		&sdr_rw_load_jump_mgr_regs->load_jump_add1);
	writel(RW_MGR_INIT_RESET_0_CKE_0,
		&sdr_rw_load_jump_mgr_regs->load_jump_add2);

	/* Execute count instruction */
	writel(RW_MGR_INIT_RESET_0_CKE_0, grpaddr);

	/* indicate that memory is stable */
	writel(1, &phy_mgr_cfg->reset_mem_stbl);

	/*
	 * transition the RESET to high
	 * Wait for 500us
	 */

	/*
	 * 500us @ 266MHz (3.75 ns) ~ 134000 clock cycles
	 * If a and b are the number of iteration in 2 nested loops
	 * it takes the following number of cycles to complete the operation
	 * number_of_cycles = ((2 + n) * a + 2) * b
	 * where n is the number of instruction in the inner loop
	 * One possible solution is n = 2 , a = 131 , b = 256 => a = 83,
	 * b = FF
	 */

	/* Load counters */
	writel(SKIP_DELAY_LOOP_VALUE_OR_ZERO(SEQ_TRESET_CNTR0_VAL),
	       &sdr_rw_load_mgr_regs->load_cntr0);
	writel(SKIP_DELAY_LOOP_VALUE_OR_ZERO(SEQ_TRESET_CNTR1_VAL),
	       &sdr_rw_load_mgr_regs->load_cntr1);
	writel(SKIP_DELAY_LOOP_VALUE_OR_ZERO(SEQ_TRESET_CNTR2_VAL),
	       &sdr_rw_load_mgr_regs->load_cntr2);

	/* Load jump address */
	writel(RW_MGR_INIT_RESET_1_CKE_0,
		&sdr_rw_load_jump_mgr_regs->load_jump_add0);
	writel(RW_MGR_INIT_RESET_1_CKE_0,
		&sdr_rw_load_jump_mgr_regs->load_jump_add1);
	writel(RW_MGR_INIT_RESET_1_CKE_0,
		&sdr_rw_load_jump_mgr_regs->load_jump_add2);

	writel(RW_MGR_INIT_RESET_1_CKE_0, grpaddr);

	/* bring up clock enable */

	/* tXRP < 250 ck cycles */
	delay_for_n_mem_clocks(250);

	for (r = 0; r < RW_MGR_MEM_NUMBER_OF_RANKS; r++) {
		if (param->skip_ranks[r]) {
			/* request to skip the rank */
			continue;
		}

		/* set rank */
		set_rank_and_odt_mask(r, RW_MGR_ODT_MODE_OFF);

		/*
		 * USER Use Mirror-ed commands for odd ranks if address
		 * mirrorring is on
		 */
		if ((RW_MGR_MEM_ADDRESS_MIRRORING >> r) & 0x1) {
			set_jump_as_return();
			writel(RW_MGR_MRS2_MIRR, grpaddr);
			delay_for_n_mem_clocks(4);
			set_jump_as_return();
			writel(RW_MGR_MRS3_MIRR, grpaddr);
			delay_for_n_mem_clocks(4);
			set_jump_as_return();
			writel(RW_MGR_MRS1_MIRR, grpaddr);
			delay_for_n_mem_clocks(4);
			set_jump_as_return();
			writel(RW_MGR_MRS0_DLL_RESET_MIRR, grpaddr);
		} else {
			set_jump_as_return();
			writel(RW_MGR_MRS2, grpaddr);
			delay_for_n_mem_clocks(4);
			set_jump_as_return();
			writel(RW_MGR_MRS3, grpaddr);
			delay_for_n_mem_clocks(4);
			set_jump_as_return();
			writel(RW_MGR_MRS1, grpaddr);
			set_jump_as_return();
			writel(RW_MGR_MRS0_DLL_RESET, grpaddr);
		}
		set_jump_as_return();
		writel(RW_MGR_ZQCL, grpaddr);

		/* tZQinit = tDLLK = 512 ck cycles */
		delay_for_n_mem_clocks(512);
	}
}

/*
 * At the end of calibration we have to program the user settings in, and
 * USER  hand off the memory to the user.
 */
static void rw_mgr_mem_handoff(void)
{
	uint32_t r;
	uint32_t grpaddr = SDR_PHYGRP_RWMGRGRP_ADDRESS |
			   RW_MGR_RUN_SINGLE_GROUP_OFFSET;

	debug("%s:%d\n", __func__, __LINE__);
	for (r = 0; r < RW_MGR_MEM_NUMBER_OF_RANKS; r++) {
		if (param->skip_ranks[r])
			/* request to skip the rank */
			continue;
		/* set rank */
		set_rank_and_odt_mask(r, RW_MGR_ODT_MODE_OFF);

		/* precharge all banks ... */
		writel(RW_MGR_PRECHARGE_ALL, grpaddr);

		/* load up MR settings specified by user */

		/*
		 * Use Mirror-ed commands for odd ranks if address
		 * mirrorring is on
		 */
		if ((RW_MGR_MEM_ADDRESS_MIRRORING >> r) & 0x1) {
			set_jump_as_return();
			writel(RW_MGR_MRS2_MIRR, grpaddr);
			delay_for_n_mem_clocks(4);
			set_jump_as_return();
			writel(RW_MGR_MRS3_MIRR, grpaddr);
			delay_for_n_mem_clocks(4);
			set_jump_as_return();
			writel(RW_MGR_MRS1_MIRR, grpaddr);
			delay_for_n_mem_clocks(4);
			set_jump_as_return();
			writel(RW_MGR_MRS0_USER_MIRR, grpaddr);
		} else {
			set_jump_as_return();
			writel(RW_MGR_MRS2, grpaddr);
			delay_for_n_mem_clocks(4);
			set_jump_as_return();
			writel(RW_MGR_MRS3, grpaddr);
			delay_for_n_mem_clocks(4);
			set_jump_as_return();
			writel(RW_MGR_MRS1, grpaddr);
			delay_for_n_mem_clocks(4);
			set_jump_as_return();
			writel(RW_MGR_MRS0_USER, grpaddr);
		}
		/*
		 * USER  need to wait tMOD (12CK or 15ns) time before issuing
		 * other commands, but we will have plenty of NIOS cycles before
		 * actual handoff so its okay.
		 */
	}
}

/*
 * performs a guaranteed read on the patterns we are going to use during a
 * read test to ensure memory works
 */
static uint32_t rw_mgr_mem_calibrate_read_test_patterns(uint32_t rank_bgn,
	uint32_t group, uint32_t num_tries, uint32_t *bit_chk,
	uint32_t all_ranks)
{
	uint32_t r, vg;
	uint32_t correct_mask_vg;
	uint32_t tmp_bit_chk;
	uint32_t rank_end = all_ranks ? RW_MGR_MEM_NUMBER_OF_RANKS :
		(rank_bgn + NUM_RANKS_PER_SHADOW_REG);
	uint32_t addr;
	uint32_t base_rw_mgr;

	*bit_chk = param->read_correct_mask;
	correct_mask_vg = param->read_correct_mask_vg;

	for (r = rank_bgn; r < rank_end; r++) {
		if (param->skip_ranks[r])
			/* request to skip the rank */
			continue;

		/* set rank */
		set_rank_and_odt_mask(r, RW_MGR_ODT_MODE_READ_WRITE);

		/* Load up a constant bursts of read commands */
		writel(0x20, &sdr_rw_load_mgr_regs->load_cntr0);
		writel(RW_MGR_GUARANTEED_READ,
			&sdr_rw_load_jump_mgr_regs->load_jump_add0);

		writel(0x20, &sdr_rw_load_mgr_regs->load_cntr1);
		writel(RW_MGR_GUARANTEED_READ_CONT,
			&sdr_rw_load_jump_mgr_regs->load_jump_add1);

		tmp_bit_chk = 0;
		for (vg = RW_MGR_MEM_VIRTUAL_GROUPS_PER_READ_DQS-1; ; vg--) {
			/* reset the fifos to get pointers to known state */

			writel(0, &phy_mgr_cmd->fifo_reset);
			writel(0, SDR_PHYGRP_RWMGRGRP_ADDRESS |
				  RW_MGR_RESET_READ_DATAPATH_OFFSET);

			tmp_bit_chk = tmp_bit_chk << (RW_MGR_MEM_DQ_PER_READ_DQS
				/ RW_MGR_MEM_VIRTUAL_GROUPS_PER_READ_DQS);

			addr = SDR_PHYGRP_RWMGRGRP_ADDRESS | RW_MGR_RUN_SINGLE_GROUP_OFFSET;
			writel(RW_MGR_GUARANTEED_READ, addr +
			       ((group * RW_MGR_MEM_VIRTUAL_GROUPS_PER_READ_DQS +
				vg) << 2));

			base_rw_mgr = readl(SDR_PHYGRP_RWMGRGRP_ADDRESS);
			tmp_bit_chk = tmp_bit_chk | (correct_mask_vg & (~base_rw_mgr));

			if (vg == 0)
				break;
		}
		*bit_chk &= tmp_bit_chk;
	}

	addr = SDR_PHYGRP_RWMGRGRP_ADDRESS | RW_MGR_RUN_SINGLE_GROUP_OFFSET;
	writel(RW_MGR_CLEAR_DQS_ENABLE, addr + (group << 2));

	set_rank_and_odt_mask(0, RW_MGR_ODT_MODE_OFF);
	debug_cond(DLEVEL == 1, "%s:%d test_load_patterns(%u,ALL) => (%u == %u) =>\
		   %lu\n", __func__, __LINE__, group, *bit_chk, param->read_correct_mask,
		   (long unsigned int)(*bit_chk == param->read_correct_mask));
	return *bit_chk == param->read_correct_mask;
}

static uint32_t rw_mgr_mem_calibrate_read_test_patterns_all_ranks
	(uint32_t group, uint32_t num_tries, uint32_t *bit_chk)
{
	return rw_mgr_mem_calibrate_read_test_patterns(0, group,
		num_tries, bit_chk, 1);
}

/* load up the patterns we are going to use during a read test */
static void rw_mgr_mem_calibrate_read_load_patterns(uint32_t rank_bgn,
	uint32_t all_ranks)
{
	uint32_t r;
	uint32_t rank_end = all_ranks ? RW_MGR_MEM_NUMBER_OF_RANKS :
		(rank_bgn + NUM_RANKS_PER_SHADOW_REG);

	debug("%s:%d\n", __func__, __LINE__);
	for (r = rank_bgn; r < rank_end; r++) {
		if (param->skip_ranks[r])
			/* request to skip the rank */
			continue;

		/* set rank */
		set_rank_and_odt_mask(r, RW_MGR_ODT_MODE_READ_WRITE);

		/* Load up a constant bursts */
		writel(0x20, &sdr_rw_load_mgr_regs->load_cntr0);

		writel(RW_MGR_GUARANTEED_WRITE_WAIT0,
			&sdr_rw_load_jump_mgr_regs->load_jump_add0);

		writel(0x20, &sdr_rw_load_mgr_regs->load_cntr1);

		writel(RW_MGR_GUARANTEED_WRITE_WAIT1,
			&sdr_rw_load_jump_mgr_regs->load_jump_add1);

		writel(0x04, &sdr_rw_load_mgr_regs->load_cntr2);

		writel(RW_MGR_GUARANTEED_WRITE_WAIT2,
			&sdr_rw_load_jump_mgr_regs->load_jump_add2);

		writel(0x04, &sdr_rw_load_mgr_regs->load_cntr3);

		writel(RW_MGR_GUARANTEED_WRITE_WAIT3,
			&sdr_rw_load_jump_mgr_regs->load_jump_add3);

		writel(RW_MGR_GUARANTEED_WRITE, SDR_PHYGRP_RWMGRGRP_ADDRESS |
						RW_MGR_RUN_SINGLE_GROUP_OFFSET);
	}

	set_rank_and_odt_mask(0, RW_MGR_ODT_MODE_OFF);
}

/*
 * try a read and see if it returns correct data back. has dummy reads
 * inserted into the mix used to align dqs enable. has more thorough checks
 * than the regular read test.
 */
static uint32_t rw_mgr_mem_calibrate_read_test(uint32_t rank_bgn, uint32_t group,
	uint32_t num_tries, uint32_t all_correct, uint32_t *bit_chk,
	uint32_t all_groups, uint32_t all_ranks)
{
	uint32_t r, vg;
	uint32_t correct_mask_vg;
	uint32_t tmp_bit_chk;
	uint32_t rank_end = all_ranks ? RW_MGR_MEM_NUMBER_OF_RANKS :
		(rank_bgn + NUM_RANKS_PER_SHADOW_REG);
	uint32_t addr;
	uint32_t base_rw_mgr;

	*bit_chk = param->read_correct_mask;
	correct_mask_vg = param->read_correct_mask_vg;

	uint32_t quick_read_mode = (((STATIC_CALIB_STEPS) &
		CALIB_SKIP_DELAY_SWEEPS) && ENABLE_SUPER_QUICK_CALIBRATION);

	for (r = rank_bgn; r < rank_end; r++) {
		if (param->skip_ranks[r])
			/* request to skip the rank */
			continue;

		/* set rank */
		set_rank_and_odt_mask(r, RW_MGR_ODT_MODE_READ_WRITE);

		writel(0x10, &sdr_rw_load_mgr_regs->load_cntr1);

		writel(RW_MGR_READ_B2B_WAIT1,
			&sdr_rw_load_jump_mgr_regs->load_jump_add1);

		writel(0x10, &sdr_rw_load_mgr_regs->load_cntr2);
		writel(RW_MGR_READ_B2B_WAIT2,
			&sdr_rw_load_jump_mgr_regs->load_jump_add2);

		if (quick_read_mode)
			writel(0x1, &sdr_rw_load_mgr_regs->load_cntr0);
			/* need at least two (1+1) reads to capture failures */
		else if (all_groups)
			writel(0x06, &sdr_rw_load_mgr_regs->load_cntr0);
		else
			writel(0x32, &sdr_rw_load_mgr_regs->load_cntr0);

		writel(RW_MGR_READ_B2B,
			&sdr_rw_load_jump_mgr_regs->load_jump_add0);
		if (all_groups)
			writel(RW_MGR_MEM_IF_READ_DQS_WIDTH *
			       RW_MGR_MEM_VIRTUAL_GROUPS_PER_READ_DQS - 1,
			       &sdr_rw_load_mgr_regs->load_cntr3);
		else
			writel(0x0, &sdr_rw_load_mgr_regs->load_cntr3);

		writel(RW_MGR_READ_B2B,
			&sdr_rw_load_jump_mgr_regs->load_jump_add3);

		tmp_bit_chk = 0;
		for (vg = RW_MGR_MEM_VIRTUAL_GROUPS_PER_READ_DQS-1; ; vg--) {
			/* reset the fifos to get pointers to known state */
			writel(0, &phy_mgr_cmd->fifo_reset);
			writel(0, SDR_PHYGRP_RWMGRGRP_ADDRESS |
				  RW_MGR_RESET_READ_DATAPATH_OFFSET);

			tmp_bit_chk = tmp_bit_chk << (RW_MGR_MEM_DQ_PER_READ_DQS
				/ RW_MGR_MEM_VIRTUAL_GROUPS_PER_READ_DQS);

			if (all_groups)
				addr = SDR_PHYGRP_RWMGRGRP_ADDRESS | RW_MGR_RUN_ALL_GROUPS_OFFSET;
			else
				addr = SDR_PHYGRP_RWMGRGRP_ADDRESS | RW_MGR_RUN_SINGLE_GROUP_OFFSET;

			writel(RW_MGR_READ_B2B, addr +
			       ((group * RW_MGR_MEM_VIRTUAL_GROUPS_PER_READ_DQS +
			       vg) << 2));

			base_rw_mgr = readl(SDR_PHYGRP_RWMGRGRP_ADDRESS);
			tmp_bit_chk = tmp_bit_chk | (correct_mask_vg & ~(base_rw_mgr));

			if (vg == 0)
				break;
		}
		*bit_chk &= tmp_bit_chk;
	}

	addr = SDR_PHYGRP_RWMGRGRP_ADDRESS | RW_MGR_RUN_SINGLE_GROUP_OFFSET;
	writel(RW_MGR_CLEAR_DQS_ENABLE, addr + (group << 2));

	if (all_correct) {
		set_rank_and_odt_mask(0, RW_MGR_ODT_MODE_OFF);
		debug_cond(DLEVEL == 2, "%s:%d read_test(%u,ALL,%u) =>\
			   (%u == %u) => %lu", __func__, __LINE__, group,
			   all_groups, *bit_chk, param->read_correct_mask,
			   (long unsigned int)(*bit_chk ==
			   param->read_correct_mask));
		return *bit_chk == param->read_correct_mask;
	} else	{
		set_rank_and_odt_mask(0, RW_MGR_ODT_MODE_OFF);
		debug_cond(DLEVEL == 2, "%s:%d read_test(%u,ONE,%u) =>\
			   (%u != %lu) => %lu\n", __func__, __LINE__,
			   group, all_groups, *bit_chk, (long unsigned int)0,
			   (long unsigned int)(*bit_chk != 0x00));
		return *bit_chk != 0x00;
	}
}

static uint32_t rw_mgr_mem_calibrate_read_test_all_ranks(uint32_t group,
	uint32_t num_tries, uint32_t all_correct, uint32_t *bit_chk,
	uint32_t all_groups)
{
	return rw_mgr_mem_calibrate_read_test(0, group, num_tries, all_correct,
					      bit_chk, all_groups, 1);
}

static void rw_mgr_incr_vfifo(uint32_t grp, uint32_t *v)
{
	writel(grp, &phy_mgr_cmd->inc_vfifo_hard_phy);
	(*v)++;
}

static void rw_mgr_decr_vfifo(uint32_t grp, uint32_t *v)
{
	uint32_t i;

	for (i = 0; i < VFIFO_SIZE-1; i++)
		rw_mgr_incr_vfifo(grp, v);
}

static int find_vfifo_read(uint32_t grp, uint32_t *bit_chk)
{
	uint32_t  v;
	uint32_t fail_cnt = 0;
	uint32_t test_status;

	for (v = 0; v < VFIFO_SIZE; ) {
		debug_cond(DLEVEL == 2, "%s:%d find_dqs_en_phase: vfifo %u\n",
			   __func__, __LINE__, v);
		test_status = rw_mgr_mem_calibrate_read_test_all_ranks
			(grp, 1, PASS_ONE_BIT, bit_chk, 0);
		if (!test_status) {
			fail_cnt++;

			if (fail_cnt == 2)
				break;
		}

		/* fiddle with FIFO */
		rw_mgr_incr_vfifo(grp, &v);
	}

	if (v >= VFIFO_SIZE) {
		/* no failing read found!! Something must have gone wrong */
		debug_cond(DLEVEL == 2, "%s:%d find_dqs_en_phase: vfifo failed\n",
			   __func__, __LINE__);
		return 0;
	} else {
		return v;
	}
}

static int find_working_phase(uint32_t *grp, uint32_t *bit_chk,
			      uint32_t dtaps_per_ptap, uint32_t *work_bgn,
			      uint32_t *v, uint32_t *d, uint32_t *p,
			      uint32_t *i, uint32_t *max_working_cnt)
{
	uint32_t found_begin = 0;
	uint32_t tmp_delay = 0;
	uint32_t test_status;

	for (*d = 0; *d <= dtaps_per_ptap; (*d)++, tmp_delay +=
		IO_DELAY_PER_DQS_EN_DCHAIN_TAP) {
		*work_bgn = tmp_delay;
		scc_mgr_set_dqs_en_delay_all_ranks(*grp, *d);

		for (*i = 0; *i < VFIFO_SIZE; (*i)++) {
			for (*p = 0; *p <= IO_DQS_EN_PHASE_MAX; (*p)++, *work_bgn +=
				IO_DELAY_PER_OPA_TAP) {
				scc_mgr_set_dqs_en_phase_all_ranks(*grp, *p);

				test_status =
				rw_mgr_mem_calibrate_read_test_all_ranks
				(*grp, 1, PASS_ONE_BIT, bit_chk, 0);

				if (test_status) {
					*max_working_cnt = 1;
					found_begin = 1;
					break;
				}
			}

			if (found_begin)
				break;

			if (*p > IO_DQS_EN_PHASE_MAX)
				/* fiddle with FIFO */
				rw_mgr_incr_vfifo(*grp, v);
		}

		if (found_begin)
			break;
	}

	if (*i >= VFIFO_SIZE) {
		/* cannot find working solution */
		debug_cond(DLEVEL == 2, "%s:%d find_dqs_en_phase: no vfifo/\
			   ptap/dtap\n", __func__, __LINE__);
		return 0;
	} else {
		return 1;
	}
}

static void sdr_backup_phase(uint32_t *grp, uint32_t *bit_chk,
			     uint32_t *work_bgn, uint32_t *v, uint32_t *d,
			     uint32_t *p, uint32_t *max_working_cnt)
{
	uint32_t found_begin = 0;
	uint32_t tmp_delay;

	/* Special case code for backing up a phase */
	if (*p == 0) {
		*p = IO_DQS_EN_PHASE_MAX;
		rw_mgr_decr_vfifo(*grp, v);
	} else {
		(*p)--;
	}
	tmp_delay = *work_bgn - IO_DELAY_PER_OPA_TAP;
	scc_mgr_set_dqs_en_phase_all_ranks(*grp, *p);

	for (*d = 0; *d <= IO_DQS_EN_DELAY_MAX && tmp_delay < *work_bgn;
		(*d)++, tmp_delay += IO_DELAY_PER_DQS_EN_DCHAIN_TAP) {
		scc_mgr_set_dqs_en_delay_all_ranks(*grp, *d);

		if (rw_mgr_mem_calibrate_read_test_all_ranks(*grp, 1,
							     PASS_ONE_BIT,
							     bit_chk, 0)) {
			found_begin = 1;
			*work_bgn = tmp_delay;
			break;
		}
	}

	/* We have found a working dtap before the ptap found above */
	if (found_begin == 1)
		(*max_working_cnt)++;

	/*
	 * Restore VFIFO to old state before we decremented it
	 * (if needed).
	 */
	(*p)++;
	if (*p > IO_DQS_EN_PHASE_MAX) {
		*p = 0;
		rw_mgr_incr_vfifo(*grp, v);
	}

	scc_mgr_set_dqs_en_delay_all_ranks(*grp, 0);
}

static int sdr_nonworking_phase(uint32_t *grp, uint32_t *bit_chk,
			     uint32_t *work_bgn, uint32_t *v, uint32_t *d,
			     uint32_t *p, uint32_t *i, uint32_t *max_working_cnt,
			     uint32_t *work_end)
{
	uint32_t found_end = 0;

	(*p)++;
	*work_end += IO_DELAY_PER_OPA_TAP;
	if (*p > IO_DQS_EN_PHASE_MAX) {
		/* fiddle with FIFO */
		*p = 0;
		rw_mgr_incr_vfifo(*grp, v);
	}

	for (; *i < VFIFO_SIZE + 1; (*i)++) {
		for (; *p <= IO_DQS_EN_PHASE_MAX; (*p)++, *work_end
			+= IO_DELAY_PER_OPA_TAP) {
			scc_mgr_set_dqs_en_phase_all_ranks(*grp, *p);

			if (!rw_mgr_mem_calibrate_read_test_all_ranks
				(*grp, 1, PASS_ONE_BIT, bit_chk, 0)) {
				found_end = 1;
				break;
			} else {
				(*max_working_cnt)++;
			}
		}

		if (found_end)
			break;

		if (*p > IO_DQS_EN_PHASE_MAX) {
			/* fiddle with FIFO */
			rw_mgr_incr_vfifo(*grp, v);
			*p = 0;
		}
	}

	if (*i >= VFIFO_SIZE + 1) {
		/* cannot see edge of failing read */
		debug_cond(DLEVEL == 2, "%s:%d sdr_nonworking_phase: end:\
			   failed\n", __func__, __LINE__);
		return 0;
	} else {
		return 1;
	}
}

static int sdr_find_window_centre(uint32_t *grp, uint32_t *bit_chk,
				  uint32_t *work_bgn, uint32_t *v, uint32_t *d,
				  uint32_t *p, uint32_t *work_mid,
				  uint32_t *work_end)
{
	int i;
	int tmp_delay = 0;

	*work_mid = (*work_bgn + *work_end) / 2;

	debug_cond(DLEVEL == 2, "work_bgn=%d work_end=%d work_mid=%d\n",
		   *work_bgn, *work_end, *work_mid);
	/* Get the middle delay to be less than a VFIFO delay */
	for (*p = 0; *p <= IO_DQS_EN_PHASE_MAX;
		(*p)++, tmp_delay += IO_DELAY_PER_OPA_TAP)
		;
	debug_cond(DLEVEL == 2, "vfifo ptap delay %d\n", tmp_delay);
	while (*work_mid > tmp_delay)
		*work_mid -= tmp_delay;
	debug_cond(DLEVEL == 2, "new work_mid %d\n", *work_mid);

	tmp_delay = 0;
	for (*p = 0; *p <= IO_DQS_EN_PHASE_MAX && tmp_delay < *work_mid;
		(*p)++, tmp_delay += IO_DELAY_PER_OPA_TAP)
		;
	tmp_delay -= IO_DELAY_PER_OPA_TAP;
	debug_cond(DLEVEL == 2, "new p %d, tmp_delay=%d\n", (*p) - 1, tmp_delay);
	for (*d = 0; *d <= IO_DQS_EN_DELAY_MAX && tmp_delay < *work_mid; (*d)++,
		tmp_delay += IO_DELAY_PER_DQS_EN_DCHAIN_TAP)
		;
	debug_cond(DLEVEL == 2, "new d %d, tmp_delay=%d\n", *d, tmp_delay);

	scc_mgr_set_dqs_en_phase_all_ranks(*grp, (*p) - 1);
	scc_mgr_set_dqs_en_delay_all_ranks(*grp, *d);

	/*
	 * push vfifo until we can successfully calibrate. We can do this
	 * because the largest possible margin in 1 VFIFO cycle.
	 */
	for (i = 0; i < VFIFO_SIZE; i++) {
		debug_cond(DLEVEL == 2, "find_dqs_en_phase: center: vfifo=%u\n",
			   *v);
		if (rw_mgr_mem_calibrate_read_test_all_ranks(*grp, 1,
							     PASS_ONE_BIT,
							     bit_chk, 0)) {
			break;
		}

		/* fiddle with FIFO */
		rw_mgr_incr_vfifo(*grp, v);
	}

	if (i >= VFIFO_SIZE) {
		debug_cond(DLEVEL == 2, "%s:%d find_dqs_en_phase: center: \
			   failed\n", __func__, __LINE__);
		return 0;
	} else {
		return 1;
	}
}

/* find a good dqs enable to use */
static uint32_t rw_mgr_mem_calibrate_vfifo_find_dqs_en_phase(uint32_t grp)
{
	uint32_t v, d, p, i;
	uint32_t max_working_cnt;
	uint32_t bit_chk;
	uint32_t dtaps_per_ptap;
	uint32_t work_bgn, work_mid, work_end;
	uint32_t found_passing_read, found_failing_read, initial_failing_dtap;

	debug("%s:%d %u\n", __func__, __LINE__, grp);

	reg_file_set_sub_stage(CAL_SUBSTAGE_VFIFO_CENTER);

	scc_mgr_set_dqs_en_delay_all_ranks(grp, 0);
	scc_mgr_set_dqs_en_phase_all_ranks(grp, 0);

	/* ************************************************************** */
	/* * Step 0 : Determine number of delay taps for each phase tap * */
	dtaps_per_ptap = IO_DELAY_PER_OPA_TAP/IO_DELAY_PER_DQS_EN_DCHAIN_TAP;

	/* ********************************************************* */
	/* * Step 1 : First push vfifo until we get a failing read * */
	v = find_vfifo_read(grp, &bit_chk);

	max_working_cnt = 0;

	/* ******************************************************** */
	/* * step 2: find first working phase, increment in ptaps * */
	work_bgn = 0;
	if (find_working_phase(&grp, &bit_chk, dtaps_per_ptap, &work_bgn, &v, &d,
				&p, &i, &max_working_cnt) == 0)
		return 0;

	work_end = work_bgn;

	/*
	 * If d is 0 then the working window covers a phase tap and
	 * we can follow the old procedure otherwise, we've found the beginning,
	 * and we need to increment the dtaps until we find the end.
	 */
	if (d == 0) {
		/* ********************************************************* */
		/* * step 3a: if we have room, back off by one and
		increment in dtaps * */

		sdr_backup_phase(&grp, &bit_chk, &work_bgn, &v, &d, &p,
				 &max_working_cnt);

		/* ********************************************************* */
		/* * step 4a: go forward from working phase to non working
		phase, increment in ptaps * */
		if (sdr_nonworking_phase(&grp, &bit_chk, &work_bgn, &v, &d, &p,
					 &i, &max_working_cnt, &work_end) == 0)
			return 0;

		/* ********************************************************* */
		/* * step 5a:  back off one from last, increment in dtaps  * */

		/* Special case code for backing up a phase */
		if (p == 0) {
			p = IO_DQS_EN_PHASE_MAX;
			rw_mgr_decr_vfifo(grp, &v);
		} else {
			p = p - 1;
		}

		work_end -= IO_DELAY_PER_OPA_TAP;
		scc_mgr_set_dqs_en_phase_all_ranks(grp, p);

		/* * The actual increment of dtaps is done outside of
		the if/else loop to share code */
		d = 0;

		debug_cond(DLEVEL == 2, "%s:%d find_dqs_en_phase: v/p: \
			   vfifo=%u ptap=%u\n", __func__, __LINE__,
			   v, p);
	} else {
		/* ******************************************************* */
		/* * step 3-5b:  Find the right edge of the window using
		delay taps   * */
		debug_cond(DLEVEL == 2, "%s:%d find_dqs_en_phase:vfifo=%u \
			   ptap=%u dtap=%u bgn=%u\n", __func__, __LINE__,
			   v, p, d, work_bgn);

		work_end = work_bgn;

		/* * The actual increment of dtaps is done outside of the
		if/else loop to share code */

		/* Only here to counterbalance a subtract later on which is
		not needed if this branch of the algorithm is taken */
		max_working_cnt++;
	}

	/* The dtap increment to find the failing edge is done here */
	for (; d <= IO_DQS_EN_DELAY_MAX; d++, work_end +=
		IO_DELAY_PER_DQS_EN_DCHAIN_TAP) {
			debug_cond(DLEVEL == 2, "%s:%d find_dqs_en_phase: \
				   end-2: dtap=%u\n", __func__, __LINE__, d);
			scc_mgr_set_dqs_en_delay_all_ranks(grp, d);

			if (!rw_mgr_mem_calibrate_read_test_all_ranks(grp, 1,
								      PASS_ONE_BIT,
								      &bit_chk, 0)) {
				break;
			}
	}

	/* Go back to working dtap */
	if (d != 0)
		work_end -= IO_DELAY_PER_DQS_EN_DCHAIN_TAP;

	debug_cond(DLEVEL == 2, "%s:%d find_dqs_en_phase: v/p/d: vfifo=%u \
		   ptap=%u dtap=%u end=%u\n", __func__, __LINE__,
		   v, p, d-1, work_end);

	if (work_end < work_bgn) {
		/* nil range */
		debug_cond(DLEVEL == 2, "%s:%d find_dqs_en_phase: end-2: \
			   failed\n", __func__, __LINE__);
		return 0;
	}

	debug_cond(DLEVEL == 2, "%s:%d find_dqs_en_phase: found range [%u,%u]\n",
		   __func__, __LINE__, work_bgn, work_end);

	/* *************************************************************** */
	/*
	 * * We need to calculate the number of dtaps that equal a ptap
	 * * To do that we'll back up a ptap and re-find the edge of the
	 * * window using dtaps
	 */

	debug_cond(DLEVEL == 2, "%s:%d find_dqs_en_phase: calculate dtaps_per_ptap \
		   for tracking\n", __func__, __LINE__);

	/* Special case code for backing up a phase */
	if (p == 0) {
		p = IO_DQS_EN_PHASE_MAX;
		rw_mgr_decr_vfifo(grp, &v);
		debug_cond(DLEVEL == 2, "%s:%d find_dqs_en_phase: backedup \
			   cycle/phase: v=%u p=%u\n", __func__, __LINE__,
			   v, p);
	} else {
		p = p - 1;
		debug_cond(DLEVEL == 2, "%s:%d find_dqs_en_phase: backedup \
			   phase only: v=%u p=%u", __func__, __LINE__,
			   v, p);
	}

	scc_mgr_set_dqs_en_phase_all_ranks(grp, p);

	/*
	 * Increase dtap until we first see a passing read (in case the
	 * window is smaller than a ptap),
	 * and then a failing read to mark the edge of the window again
	 */

	/* Find a passing read */
	debug_cond(DLEVEL == 2, "%s:%d find_dqs_en_phase: find passing read\n",
		   __func__, __LINE__);
	found_passing_read = 0;
	found_failing_read = 0;
	initial_failing_dtap = d;
	for (; d <= IO_DQS_EN_DELAY_MAX; d++) {
		debug_cond(DLEVEL == 2, "%s:%d find_dqs_en_phase: testing \
			   read d=%u\n", __func__, __LINE__, d);
		scc_mgr_set_dqs_en_delay_all_ranks(grp, d);

		if (rw_mgr_mem_calibrate_read_test_all_ranks(grp, 1,
							     PASS_ONE_BIT,
							     &bit_chk, 0)) {
			found_passing_read = 1;
			break;
		}
	}

	if (found_passing_read) {
		/* Find a failing read */
		debug_cond(DLEVEL == 2, "%s:%d find_dqs_en_phase: find failing \
			   read\n", __func__, __LINE__);
		for (d = d + 1; d <= IO_DQS_EN_DELAY_MAX; d++) {
			debug_cond(DLEVEL == 2, "%s:%d find_dqs_en_phase: \
				   testing read d=%u\n", __func__, __LINE__, d);
			scc_mgr_set_dqs_en_delay_all_ranks(grp, d);

			if (!rw_mgr_mem_calibrate_read_test_all_ranks
				(grp, 1, PASS_ONE_BIT, &bit_chk, 0)) {
				found_failing_read = 1;
				break;
			}
		}
	} else {
		debug_cond(DLEVEL == 1, "%s:%d find_dqs_en_phase: failed to \
			   calculate dtaps", __func__, __LINE__);
		debug_cond(DLEVEL == 1, "per ptap. Fall back on static value\n");
	}

	/*
	 * The dynamically calculated dtaps_per_ptap is only valid if we
	 * found a passing/failing read. If we didn't, it means d hit the max
	 * (IO_DQS_EN_DELAY_MAX). Otherwise, dtaps_per_ptap retains its
	 * statically calculated value.
	 */
	if (found_passing_read && found_failing_read)
		dtaps_per_ptap = d - initial_failing_dtap;

	writel(dtaps_per_ptap, &sdr_reg_file->dtaps_per_ptap);
	debug_cond(DLEVEL == 2, "%s:%d find_dqs_en_phase: dtaps_per_ptap=%u \
		   - %u = %u",  __func__, __LINE__, d,
		   initial_failing_dtap, dtaps_per_ptap);

	/* ******************************************** */
	/* * step 6:  Find the centre of the window   * */
	if (sdr_find_window_centre(&grp, &bit_chk, &work_bgn, &v, &d, &p,
				   &work_mid, &work_end) == 0)
		return 0;

	debug_cond(DLEVEL == 2, "%s:%d find_dqs_en_phase: center found: \
		   vfifo=%u ptap=%u dtap=%u\n", __func__, __LINE__,
		   v, p-1, d);
	return 1;
}

/*
 * Try rw_mgr_mem_calibrate_vfifo_find_dqs_en_phase across different
 * dq_in_delay values
 */
static uint32_t
rw_mgr_mem_calibrate_vfifo_find_dqs_en_phase_sweep_dq_in_delay
(uint32_t write_group, uint32_t read_group, uint32_t test_bgn)
{
	uint32_t found;
	uint32_t i;
	uint32_t p;
	uint32_t d;
	uint32_t r;

	const uint32_t delay_step = IO_IO_IN_DELAY_MAX /
		(RW_MGR_MEM_DQ_PER_READ_DQS-1);
		/* we start at zero, so have one less dq to devide among */

	debug("%s:%d (%u,%u,%u)", __func__, __LINE__, write_group, read_group,
	      test_bgn);

	/* try different dq_in_delays since the dq path is shorter than dqs */

	for (r = 0; r < RW_MGR_MEM_NUMBER_OF_RANKS;
	     r += NUM_RANKS_PER_SHADOW_REG) {
		for (i = 0, p = test_bgn, d = 0; i < RW_MGR_MEM_DQ_PER_READ_DQS; i++, p++, d += delay_step) {
			debug_cond(DLEVEL == 1, "%s:%d rw_mgr_mem_calibrate_\
				   vfifo_find_dqs_", __func__, __LINE__);
			debug_cond(DLEVEL == 1, "en_phase_sweep_dq_in_delay: g=%u/%u ",
			       write_group, read_group);
			debug_cond(DLEVEL == 1, "r=%u, i=%u p=%u d=%u\n", r, i , p, d);
			scc_mgr_set_dq_in_delay(p, d);
			scc_mgr_load_dq(p);
		}
		writel(0, &sdr_scc_mgr->update);
	}

	found = rw_mgr_mem_calibrate_vfifo_find_dqs_en_phase(read_group);

	debug_cond(DLEVEL == 1, "%s:%d rw_mgr_mem_calibrate_vfifo_find_dqs_\
		   en_phase_sweep_dq", __func__, __LINE__);
	debug_cond(DLEVEL == 1, "_in_delay: g=%u/%u found=%u; Reseting delay \
		   chain to zero\n", write_group, read_group, found);

	for (r = 0; r < RW_MGR_MEM_NUMBER_OF_RANKS;
	     r += NUM_RANKS_PER_SHADOW_REG) {
		for (i = 0, p = test_bgn; i < RW_MGR_MEM_DQ_PER_READ_DQS;
			i++, p++) {
			scc_mgr_set_dq_in_delay(p, 0);
			scc_mgr_load_dq(p);
		}
		writel(0, &sdr_scc_mgr->update);
	}

	return found;
}

/* per-bit deskew DQ and center */
static uint32_t rw_mgr_mem_calibrate_vfifo_center(uint32_t rank_bgn,
	uint32_t write_group, uint32_t read_group, uint32_t test_bgn,
	uint32_t use_read_test, uint32_t update_fom)
{
	uint32_t i, p, d, min_index;
	/*
	 * Store these as signed since there are comparisons with
	 * signed numbers.
	 */
	uint32_t bit_chk;
	uint32_t sticky_bit_chk;
	int32_t left_edge[RW_MGR_MEM_DQ_PER_READ_DQS];
	int32_t right_edge[RW_MGR_MEM_DQ_PER_READ_DQS];
	int32_t final_dq[RW_MGR_MEM_DQ_PER_READ_DQS];
	int32_t mid;
	int32_t orig_mid_min, mid_min;
	int32_t new_dqs, start_dqs, start_dqs_en, shift_dq, final_dqs,
		final_dqs_en;
	int32_t dq_margin, dqs_margin;
	uint32_t stop;
	uint32_t temp_dq_in_delay1, temp_dq_in_delay2;
	uint32_t addr;

	debug("%s:%d: %u %u", __func__, __LINE__, read_group, test_bgn);

	addr = SDR_PHYGRP_SCCGRP_ADDRESS | SCC_MGR_DQS_IN_DELAY_OFFSET;
	start_dqs = readl(addr + (read_group << 2));
	if (IO_SHIFT_DQS_EN_WHEN_SHIFT_DQS)
		start_dqs_en = readl(addr + ((read_group << 2)
				     - IO_DQS_EN_DELAY_OFFSET));

	/* set the left and right edge of each bit to an illegal value */
	/* use (IO_IO_IN_DELAY_MAX + 1) as an illegal value */
	sticky_bit_chk = 0;
	for (i = 0; i < RW_MGR_MEM_DQ_PER_READ_DQS; i++) {
		left_edge[i]  = IO_IO_IN_DELAY_MAX + 1;
		right_edge[i] = IO_IO_IN_DELAY_MAX + 1;
	}

	/* Search for the left edge of the window for each bit */
	for (d = 0; d <= IO_IO_IN_DELAY_MAX; d++) {
		scc_mgr_apply_group_dq_in_delay(write_group, test_bgn, d);

		writel(0, &sdr_scc_mgr->update);

		/*
		 * Stop searching when the read test doesn't pass AND when
		 * we've seen a passing read on every bit.
		 */
		if (use_read_test) {
			stop = !rw_mgr_mem_calibrate_read_test(rank_bgn,
				read_group, NUM_READ_PB_TESTS, PASS_ONE_BIT,
				&bit_chk, 0, 0);
		} else {
			rw_mgr_mem_calibrate_write_test(rank_bgn, write_group,
							0, PASS_ONE_BIT,
							&bit_chk, 0);
			bit_chk = bit_chk >> (RW_MGR_MEM_DQ_PER_READ_DQS *
				(read_group - (write_group *
					RW_MGR_MEM_IF_READ_DQS_WIDTH /
					RW_MGR_MEM_IF_WRITE_DQS_WIDTH)));
			stop = (bit_chk == 0);
		}
		sticky_bit_chk = sticky_bit_chk | bit_chk;
		stop = stop && (sticky_bit_chk == param->read_correct_mask);
		debug_cond(DLEVEL == 2, "%s:%d vfifo_center(left): dtap=%u => %u == %u \
			   && %u", __func__, __LINE__, d,
			   sticky_bit_chk,
			param->read_correct_mask, stop);

		if (stop == 1) {
			break;
		} else {
			for (i = 0; i < RW_MGR_MEM_DQ_PER_READ_DQS; i++) {
				if (bit_chk & 1) {
					/* Remember a passing test as the
					left_edge */
					left_edge[i] = d;
				} else {
					/* If a left edge has not been seen yet,
					then a future passing test will mark
					this edge as the right edge */
					if (left_edge[i] ==
						IO_IO_IN_DELAY_MAX + 1) {
						right_edge[i] = -(d + 1);
					}
				}
				bit_chk = bit_chk >> 1;
			}
		}
	}

	/* Reset DQ delay chains to 0 */
	scc_mgr_apply_group_dq_in_delay(test_bgn, 0);
	sticky_bit_chk = 0;
	for (i = RW_MGR_MEM_DQ_PER_READ_DQS - 1;; i--) {
		debug_cond(DLEVEL == 2, "%s:%d vfifo_center: left_edge[%u]: \
			   %d right_edge[%u]: %d\n", __func__, __LINE__,
			   i, left_edge[i], i, right_edge[i]);

		/*
		 * Check for cases where we haven't found the left edge,
		 * which makes our assignment of the the right edge invalid.
		 * Reset it to the illegal value.
		 */
		if ((left_edge[i] == IO_IO_IN_DELAY_MAX + 1) && (
			right_edge[i] != IO_IO_IN_DELAY_MAX + 1)) {
			right_edge[i] = IO_IO_IN_DELAY_MAX + 1;
			debug_cond(DLEVEL == 2, "%s:%d vfifo_center: reset \
				   right_edge[%u]: %d\n", __func__, __LINE__,
				   i, right_edge[i]);
		}

		/*
		 * Reset sticky bit (except for bits where we have seen
		 * both the left and right edge).
		 */
		sticky_bit_chk = sticky_bit_chk << 1;
		if ((left_edge[i] != IO_IO_IN_DELAY_MAX + 1) &&
		    (right_edge[i] != IO_IO_IN_DELAY_MAX + 1)) {
			sticky_bit_chk = sticky_bit_chk | 1;
		}

		if (i == 0)
			break;
	}

	/* Search for the right edge of the window for each bit */
	for (d = 0; d <= IO_DQS_IN_DELAY_MAX - start_dqs; d++) {
		scc_mgr_set_dqs_bus_in_delay(read_group, d + start_dqs);
		if (IO_SHIFT_DQS_EN_WHEN_SHIFT_DQS) {
			uint32_t delay = d + start_dqs_en;
			if (delay > IO_DQS_EN_DELAY_MAX)
				delay = IO_DQS_EN_DELAY_MAX;
			scc_mgr_set_dqs_en_delay(read_group, delay);
		}
		scc_mgr_load_dqs(read_group);

		writel(0, &sdr_scc_mgr->update);

		/*
		 * Stop searching when the read test doesn't pass AND when
		 * we've seen a passing read on every bit.
		 */
		if (use_read_test) {
			stop = !rw_mgr_mem_calibrate_read_test(rank_bgn,
				read_group, NUM_READ_PB_TESTS, PASS_ONE_BIT,
				&bit_chk, 0, 0);
		} else {
			rw_mgr_mem_calibrate_write_test(rank_bgn, write_group,
							0, PASS_ONE_BIT,
							&bit_chk, 0);
			bit_chk = bit_chk >> (RW_MGR_MEM_DQ_PER_READ_DQS *
				(read_group - (write_group *
					RW_MGR_MEM_IF_READ_DQS_WIDTH /
					RW_MGR_MEM_IF_WRITE_DQS_WIDTH)));
			stop = (bit_chk == 0);
		}
		sticky_bit_chk = sticky_bit_chk | bit_chk;
		stop = stop && (sticky_bit_chk == param->read_correct_mask);

		debug_cond(DLEVEL == 2, "%s:%d vfifo_center(right): dtap=%u => %u == \
			   %u && %u", __func__, __LINE__, d,
			   sticky_bit_chk, param->read_correct_mask, stop);

		if (stop == 1) {
			break;
		} else {
			for (i = 0; i < RW_MGR_MEM_DQ_PER_READ_DQS; i++) {
				if (bit_chk & 1) {
					/* Remember a passing test as
					the right_edge */
					right_edge[i] = d;
				} else {
					if (d != 0) {
						/* If a right edge has not been
						seen yet, then a future passing
						test will mark this edge as the
						left edge */
						if (right_edge[i] ==
						IO_IO_IN_DELAY_MAX + 1) {
							left_edge[i] = -(d + 1);
						}
					} else {
						/* d = 0 failed, but it passed
						when testing the left edge,
						so it must be marginal,
						set it to -1 */
						if (right_edge[i] ==
							IO_IO_IN_DELAY_MAX + 1 &&
							left_edge[i] !=
							IO_IO_IN_DELAY_MAX
							+ 1) {
							right_edge[i] = -1;
						}
						/* If a right edge has not been
						seen yet, then a future passing
						test will mark this edge as the
						left edge */
						else if (right_edge[i] ==
							IO_IO_IN_DELAY_MAX +
							1) {
							left_edge[i] = -(d + 1);
						}
					}
				}

				debug_cond(DLEVEL == 2, "%s:%d vfifo_center[r,\
					   d=%u]: ", __func__, __LINE__, d);
				debug_cond(DLEVEL == 2, "bit_chk_test=%d left_edge[%u]: %d ",
					   (int)(bit_chk & 1), i, left_edge[i]);
				debug_cond(DLEVEL == 2, "right_edge[%u]: %d\n", i,
					   right_edge[i]);
				bit_chk = bit_chk >> 1;
			}
		}
	}

	/* Check that all bits have a window */
	for (i = 0; i < RW_MGR_MEM_DQ_PER_READ_DQS; i++) {
		debug_cond(DLEVEL == 2, "%s:%d vfifo_center: left_edge[%u]: \
			   %d right_edge[%u]: %d", __func__, __LINE__,
			   i, left_edge[i], i, right_edge[i]);
		if ((left_edge[i] == IO_IO_IN_DELAY_MAX + 1) || (right_edge[i]
			== IO_IO_IN_DELAY_MAX + 1)) {
			/*
			 * Restore delay chain settings before letting the loop
			 * in rw_mgr_mem_calibrate_vfifo to retry different
			 * dqs/ck relationships.
			 */
			scc_mgr_set_dqs_bus_in_delay(read_group, start_dqs);
			if (IO_SHIFT_DQS_EN_WHEN_SHIFT_DQS) {
				scc_mgr_set_dqs_en_delay(read_group,
							 start_dqs_en);
			}
			scc_mgr_load_dqs(read_group);
			writel(0, &sdr_scc_mgr->update);

			debug_cond(DLEVEL == 1, "%s:%d vfifo_center: failed to \
				   find edge [%u]: %d %d", __func__, __LINE__,
				   i, left_edge[i], right_edge[i]);
			if (use_read_test) {
				set_failing_group_stage(read_group *
					RW_MGR_MEM_DQ_PER_READ_DQS + i,
					CAL_STAGE_VFIFO,
					CAL_SUBSTAGE_VFIFO_CENTER);
			} else {
				set_failing_group_stage(read_group *
					RW_MGR_MEM_DQ_PER_READ_DQS + i,
					CAL_STAGE_VFIFO_AFTER_WRITES,
					CAL_SUBSTAGE_VFIFO_CENTER);
			}
			return 0;
		}
	}

	/* Find middle of window for each DQ bit */
	mid_min = left_edge[0] - right_edge[0];
	min_index = 0;
	for (i = 1; i < RW_MGR_MEM_DQ_PER_READ_DQS; i++) {
		mid = left_edge[i] - right_edge[i];
		if (mid < mid_min) {
			mid_min = mid;
			min_index = i;
		}
	}

	/*
	 * -mid_min/2 represents the amount that we need to move DQS.
	 * If mid_min is odd and positive we'll need to add one to
	 * make sure the rounding in further calculations is correct
	 * (always bias to the right), so just add 1 for all positive values.
	 */
	if (mid_min > 0)
		mid_min++;

	mid_min = mid_min / 2;

	debug_cond(DLEVEL == 1, "%s:%d vfifo_center: mid_min=%d (index=%u)\n",
		   __func__, __LINE__, mid_min, min_index);

	/* Determine the amount we can change DQS (which is -mid_min) */
	orig_mid_min = mid_min;
	new_dqs = start_dqs - mid_min;
	if (new_dqs > IO_DQS_IN_DELAY_MAX)
		new_dqs = IO_DQS_IN_DELAY_MAX;
	else if (new_dqs < 0)
		new_dqs = 0;

	mid_min = start_dqs - new_dqs;
	debug_cond(DLEVEL == 1, "vfifo_center: new mid_min=%d new_dqs=%d\n",
		   mid_min, new_dqs);

	if (IO_SHIFT_DQS_EN_WHEN_SHIFT_DQS) {
		if (start_dqs_en - mid_min > IO_DQS_EN_DELAY_MAX)
			mid_min += start_dqs_en - mid_min - IO_DQS_EN_DELAY_MAX;
		else if (start_dqs_en - mid_min < 0)
			mid_min += start_dqs_en - mid_min;
	}
	new_dqs = start_dqs - mid_min;

	debug_cond(DLEVEL == 1, "vfifo_center: start_dqs=%d start_dqs_en=%d \
		   new_dqs=%d mid_min=%d\n", start_dqs,
		   IO_SHIFT_DQS_EN_WHEN_SHIFT_DQS ? start_dqs_en : -1,
		   new_dqs, mid_min);

	/* Initialize data for export structures */
	dqs_margin = IO_IO_IN_DELAY_MAX + 1;
	dq_margin  = IO_IO_IN_DELAY_MAX + 1;

	/* add delay to bring centre of all DQ windows to the same "level" */
	for (i = 0, p = test_bgn; i < RW_MGR_MEM_DQ_PER_READ_DQS; i++, p++) {
		/* Use values before divide by 2 to reduce round off error */
		shift_dq = (left_edge[i] - right_edge[i] -
			(left_edge[min_index] - right_edge[min_index]))/2  +
			(orig_mid_min - mid_min);

		debug_cond(DLEVEL == 2, "vfifo_center: before: \
			   shift_dq[%u]=%d\n", i, shift_dq);

		addr = SDR_PHYGRP_SCCGRP_ADDRESS | SCC_MGR_IO_IN_DELAY_OFFSET;
		temp_dq_in_delay1 = readl(addr + (p << 2));
		temp_dq_in_delay2 = readl(addr + (i << 2));

		if (shift_dq + (int32_t)temp_dq_in_delay1 >
			(int32_t)IO_IO_IN_DELAY_MAX) {
			shift_dq = (int32_t)IO_IO_IN_DELAY_MAX - temp_dq_in_delay2;
		} else if (shift_dq + (int32_t)temp_dq_in_delay1 < 0) {
			shift_dq = -(int32_t)temp_dq_in_delay1;
		}
		debug_cond(DLEVEL == 2, "vfifo_center: after: \
			   shift_dq[%u]=%d\n", i, shift_dq);
		final_dq[i] = temp_dq_in_delay1 + shift_dq;
		scc_mgr_set_dq_in_delay(p, final_dq[i]);
		scc_mgr_load_dq(p);

		debug_cond(DLEVEL == 2, "vfifo_center: margin[%u]=[%d,%d]\n", i,
			   left_edge[i] - shift_dq + (-mid_min),
			   right_edge[i] + shift_dq - (-mid_min));
		/* To determine values for export structures */
		if (left_edge[i] - shift_dq + (-mid_min) < dq_margin)
			dq_margin = left_edge[i] - shift_dq + (-mid_min);

		if (right_edge[i] + shift_dq - (-mid_min) < dqs_margin)
			dqs_margin = right_edge[i] + shift_dq - (-mid_min);
	}

	final_dqs = new_dqs;
	if (IO_SHIFT_DQS_EN_WHEN_SHIFT_DQS)
		final_dqs_en = start_dqs_en - mid_min;

	/* Move DQS-en */
	if (IO_SHIFT_DQS_EN_WHEN_SHIFT_DQS) {
		scc_mgr_set_dqs_en_delay(read_group, final_dqs_en);
		scc_mgr_load_dqs(read_group);
	}

	/* Move DQS */
	scc_mgr_set_dqs_bus_in_delay(read_group, final_dqs);
	scc_mgr_load_dqs(read_group);
	debug_cond(DLEVEL == 2, "%s:%d vfifo_center: dq_margin=%d \
		   dqs_margin=%d", __func__, __LINE__,
		   dq_margin, dqs_margin);

	/*
	 * Do not remove this line as it makes sure all of our decisions
	 * have been applied. Apply the update bit.
	 */
	writel(0, &sdr_scc_mgr->update);

	return (dq_margin >= 0) && (dqs_margin >= 0);
}

/*
 * calibrate the read valid prediction FIFO.
 *
 *  - read valid prediction will consist of finding a good DQS enable phase,
 * DQS enable delay, DQS input phase, and DQS input delay.
 *  - we also do a per-bit deskew on the DQ lines.