sequencer.c

					 * testing the left edge, so it must be
					 * marginal, set it to -1
					 */
					if (right_edge[i] == delay_max + 1 &&
					    left_edge[i] != delay_max + 1)
						right_edge[i] = -1;
				}
			}
			break;
		}

		/* stop != 1 */
		for (i = 0; i < per_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] == 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] == delay_max + 1 &&
					    left_edge[i] != 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] == delay_max + 1)
						left_edge[i] = -(d + 1);
				}
			}

			debug_cond(DLEVEL == 2, "%s:%d center[r,d=%u]: ",
				   __func__, __LINE__, d);
			debug_cond(DLEVEL == 2,
				   "bit_chk_test=%i left_edge[%u]: %d ",
				   *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 < per_dqs; i++) {
		debug_cond(DLEVEL == 2,
			   "%s:%d write_center: left_edge[%u]: %d right_edge[%u]: %d",
			   __func__, __LINE__, i, left_edge[i],
			   i, right_edge[i]);
		if ((left_edge[i] == dqs_max + 1) ||
		    (right_edge[i] == dqs_max + 1))
			return i + 1;	/* FIXME: If we fail, retval > 0 */
	}

	return 0;
}

/* 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, 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 temp_dq_in_delay1, temp_dq_in_delay2;
	uint32_t addr;
	int ret;

	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 */
	search_left_edge(0, rank_bgn, write_group, read_group, test_bgn,
			 &bit_chk, &sticky_bit_chk,
			 left_edge, right_edge, use_read_test);

	/* Search for the right edge of the window for each bit */
	ret = search_right_edge(0, rank_bgn, write_group, read_group,
				start_dqs, start_dqs_en,
				&bit_chk, &sticky_bit_chk,
				left_edge, right_edge, use_read_test);
	if (ret) {
		/*
		 * 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);
}

/**
 * rw_mgr_mem_calibrate_guaranteed_write() - Perform guaranteed write into the device
 * @rw_group:	Read/Write Group
 * @phase:	DQ/DQS phase
 *
 * Because initially no communication ca be reliably performed with the memory
 * device, the sequencer uses a guaranteed write mechanism to write data into
 * the memory device.
 */
static int rw_mgr_mem_calibrate_guaranteed_write(const u32 rw_group,
						 const u32 phase)
{
	int ret;

	/* Set a particular DQ/DQS phase. */
	scc_mgr_set_dqdqs_output_phase_all_ranks(rw_group, phase);

	debug_cond(DLEVEL == 1, "%s:%d guaranteed write: g=%u p=%u\n",
		   __func__, __LINE__, rw_group, phase);

	/*
	 * Altera EMI_RM 2015.05.04 :: Figure 1-25
	 * Load up the patterns used by read calibration using the
	 * current DQDQS phase.
	 */
	rw_mgr_mem_calibrate_read_load_patterns(0, 1);

	if (gbl->phy_debug_mode_flags & PHY_DEBUG_DISABLE_GUARANTEED_READ)
		return 0;

	/*
	 * Altera EMI_RM 2015.05.04 :: Figure 1-26
	 * Back-to-Back reads of the patterns used for calibration.
	 */
	ret = rw_mgr_mem_calibrate_read_test_patterns(0, rw_group, 1);
	if (ret)
		debug_cond(DLEVEL == 1,
			   "%s:%d Guaranteed read test failed: g=%u p=%u\n",
			   __func__, __LINE__, rw_group, phase);
	return ret;
}

/**
 * rw_mgr_mem_calibrate_dqs_enable_calibration() - DQS Enable Calibration
 * @rw_group:	Read/Write Group
 * @test_bgn:	Rank at which the test begins
 *
 * DQS enable calibration ensures reliable capture of the DQ signal without
 * glitches on the DQS line.
 */
static int rw_mgr_mem_calibrate_dqs_enable_calibration(const u32 rw_group,
						       const u32 test_bgn)
{
	/*
	 * Altera EMI_RM 2015.05.04 :: Figure 1-27
	 * DQS and DQS Eanble Signal Relationships.
	 */

	/* We start at zero, so have one less dq to devide among */
	const u32 delay_step = IO_IO_IN_DELAY_MAX /
			       (RW_MGR_MEM_DQ_PER_READ_DQS - 1);
	int ret;
	u32 i, p, d, r;

	debug("%s:%d (%u,%u)\n", __func__, __LINE__, rw_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: g=%u r=%u i=%u p=%u d=%u\n",
				   __func__, __LINE__, rw_group, r, i, p, d);

			scc_mgr_set_dq_in_delay(p, d);
			scc_mgr_load_dq(p);
		}

		writel(0, &sdr_scc_mgr->update);
	}

	/*
	 * Try rw_mgr_mem_calibrate_vfifo_find_dqs_en_phase across different
	 * dq_in_delay values
	 */
	ret = rw_mgr_mem_calibrate_vfifo_find_dqs_en_phase(rw_group);

	debug_cond(DLEVEL == 1,
		   "%s:%d: g=%u found=%u; Reseting delay chain to zero\n",
		   __func__, __LINE__, rw_group, !ret);

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

	return ret;
}

/**
 * rw_mgr_mem_calibrate_dq_dqs_centering() - Centering DQ/DQS
 * @rw_group:		Read/Write Group
 * @test_bgn:		Rank at which the test begins
 * @use_read_test:	Perform a read test
 * @update_fom:		Update FOM
 *
 * The centerin DQ/DQS stage attempts to align DQ and DQS signals on reads
 * within a group.
 */
static int
rw_mgr_mem_calibrate_dq_dqs_centering(const u32 rw_group, const u32 test_bgn,
				      const int use_read_test,
				      const int update_fom)

{
	int ret, grp_calibrated;
	u32 rank_bgn, sr;

	/*
	 * Altera EMI_RM 2015.05.04 :: Figure 1-28
	 * Read per-bit deskew can be done on a per shadow register basis.
	 */
	grp_calibrated = 1;
	for (rank_bgn = 0, sr = 0;
	     rank_bgn < RW_MGR_MEM_NUMBER_OF_RANKS;
	     rank_bgn += NUM_RANKS_PER_SHADOW_REG, sr++) {
		/* Check if this set of ranks should be skipped entirely. */
		if (param->skip_shadow_regs[sr])
			continue;

		ret = rw_mgr_mem_calibrate_vfifo_center(rank_bgn, rw_group,
							rw_group, test_bgn,
							use_read_test,
							update_fom);
		if (ret)
			continue;

		grp_calibrated = 0;
	}

	if (!grp_calibrated)
		return -EIO;

	return 0;
}

/**
 * rw_mgr_mem_calibrate_vfifo() - Calibrate the read valid prediction FIFO
 * @rw_group:		Read/Write Group
 * @test_bgn:		Rank at which the test begins
 *
 * Stage 1: Calibrate the read valid prediction FIFO.
 *
 * This function implements UniPHY calibration Stage 1, as explained in
 * detail in Altera EMI_RM 2015.05.04 , "UniPHY Calibration Stages".
 *
 * - read valid prediction will consist of finding:
 *   - DQS enable phase and DQS enable delay (DQS Enable Calibration)
 *   - DQS input phase  and DQS input delay (DQ/DQS Centering)
 *  - we also do a per-bit deskew on the DQ lines.
 */
static int rw_mgr_mem_calibrate_vfifo(const u32 rw_group, const u32 test_bgn)
{
	uint32_t p, d;
	uint32_t dtaps_per_ptap;
	uint32_t failed_substage;

	int ret;

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

	/* Update info for sims */
	reg_file_set_group(rw_group);
	reg_file_set_stage(CAL_STAGE_VFIFO);
	reg_file_set_sub_stage(CAL_SUBSTAGE_GUARANTEED_READ);

	failed_substage = CAL_SUBSTAGE_GUARANTEED_READ;

	/* USER Determine number of delay taps for each phase tap. */
	dtaps_per_ptap = DIV_ROUND_UP(IO_DELAY_PER_OPA_TAP,
				      IO_DELAY_PER_DQS_EN_DCHAIN_TAP) - 1;

	for (d = 0; d <= dtaps_per_ptap; d += 2) {
		/*
		 * In RLDRAMX we may be messing the delay of pins in
		 * the same write rw_group but outside of the current read
		 * the rw_group, but that's ok because we haven't calibrated
		 * output side yet.
		 */
		if (d > 0) {
			scc_mgr_apply_group_all_out_delay_add_all_ranks(
								rw_group, d);
		}

		for (p = 0; p <= IO_DQDQS_OUT_PHASE_MAX; p++) {
			/* 1) Guaranteed Write */
			ret = rw_mgr_mem_calibrate_guaranteed_write(rw_group, p);
			if (ret)
				break;

			/* 2) DQS Enable Calibration */
			ret = rw_mgr_mem_calibrate_dqs_enable_calibration(rw_group,
									  test_bgn);
			if (ret) {
				failed_substage = CAL_SUBSTAGE_DQS_EN_PHASE;
				continue;
			}

			/* 3) Centering DQ/DQS */
			/*
			 * If doing read after write calibration, do not update
			 * FOM now. Do it then.
			 */
			ret = rw_mgr_mem_calibrate_dq_dqs_centering(rw_group,
								test_bgn, 1, 0);
			if (ret) {
				failed_substage = CAL_SUBSTAGE_VFIFO_CENTER;
				continue;
			}

			/* All done. */
			goto cal_done_ok;
		}
	}

	/* Calibration Stage 1 failed. */
	set_failing_group_stage(rw_group, CAL_STAGE_VFIFO, failed_substage);
	return 0;

	/* Calibration Stage 1 completed OK. */
cal_done_ok:
	/*
	 * Reset the delay chains back to zero if they have moved > 1
	 * (check for > 1 because loop will increase d even when pass in
	 * first case).
	 */
	if (d > 2)
		scc_mgr_zero_group(rw_group, 1);

	return 1;
}

/* VFIFO Calibration -- Read Deskew Calibration after write deskew */
static uint32_t rw_mgr_mem_calibrate_vfifo_end(uint32_t read_group,
					       uint32_t test_bgn)
{
	uint32_t rank_bgn, sr;
	uint32_t grp_calibrated;
	uint32_t write_group;

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

	/* update info for sims */

	reg_file_set_stage(CAL_STAGE_VFIFO_AFTER_WRITES);
	reg_file_set_sub_stage(CAL_SUBSTAGE_VFIFO_CENTER);

	write_group = read_group;

	/* update info for sims */
	reg_file_set_group(read_group);

	grp_calibrated = 1;
	/* Read per-bit deskew can be done on a per shadow register basis */
	for (rank_bgn = 0, sr = 0; rank_bgn < RW_MGR_MEM_NUMBER_OF_RANKS;
		rank_bgn += NUM_RANKS_PER_SHADOW_REG, ++sr) {
		/* Determine if this set of ranks should be skipped entirely */
		if (!param->skip_shadow_regs[sr]) {
		/* This is the last calibration round, update FOM here */
			if (!rw_mgr_mem_calibrate_vfifo_center(rank_bgn,
								write_group,
								read_group,
								test_bgn, 0,
								1)) {
				grp_calibrated = 0;
			}
		}
	}


	if (grp_calibrated == 0) {
		set_failing_group_stage(write_group,
					CAL_STAGE_VFIFO_AFTER_WRITES,
					CAL_SUBSTAGE_VFIFO_CENTER);
		return 0;
	}

	return 1;
}

/* Calibrate LFIFO to find smallest read latency */
static uint32_t rw_mgr_mem_calibrate_lfifo(void)
{
	uint32_t found_one;

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

	/* update info for sims */
	reg_file_set_stage(CAL_STAGE_LFIFO);
	reg_file_set_sub_stage(CAL_SUBSTAGE_READ_LATENCY);

	/* Load up the patterns used by read calibration for all ranks */
	rw_mgr_mem_calibrate_read_load_patterns(0, 1);
	found_one = 0;

	do {
		writel(gbl->curr_read_lat, &phy_mgr_cfg->phy_rlat);
		debug_cond(DLEVEL == 2, "%s:%d lfifo: read_lat=%u",
			   __func__, __LINE__, gbl->curr_read_lat);

		if (!rw_mgr_mem_calibrate_read_test_all_ranks(0,
							      NUM_READ_TESTS,
							      PASS_ALL_BITS,
							      1)) {
			break;
		}

		found_one = 1;
		/* reduce read latency and see if things are working */
		/* correctly */
		gbl->curr_read_lat--;
	} while (gbl->curr_read_lat > 0);

	/* reset the fifos to get pointers to known state */

	writel(0, &phy_mgr_cmd->fifo_reset);

	if (found_one) {
		/* add a fudge factor to the read latency that was determined */
		gbl->curr_read_lat += 2;
		writel(gbl->curr_read_lat, &phy_mgr_cfg->phy_rlat);
		debug_cond(DLEVEL == 2, "%s:%d lfifo: success: using \
			   read_lat=%u\n", __func__, __LINE__,
			   gbl->curr_read_lat);
		return 1;
	} else {
		set_failing_group_stage(0xff, CAL_STAGE_LFIFO,
					CAL_SUBSTAGE_READ_LATENCY);

		debug_cond(DLEVEL == 2, "%s:%d lfifo: failed at initial \
			   read_lat=%u\n", __func__, __LINE__,
			   gbl->curr_read_lat);
		return 0;
	}
}

/*
 * issue write test command.
 * two variants are provided. one that just tests a write pattern and
 * another that tests datamask functionality.
 */
static void rw_mgr_mem_calibrate_write_test_issue(uint32_t group,
						  uint32_t test_dm)
{
	uint32_t mcc_instruction;
	uint32_t quick_write_mode = (((STATIC_CALIB_STEPS) & CALIB_SKIP_WRITES) &&
		ENABLE_SUPER_QUICK_CALIBRATION);
	uint32_t rw_wl_nop_cycles;
	uint32_t addr;

	/*
	 * Set counter and jump addresses for the right
	 * number of NOP cycles.
	 * The number of supported NOP cycles can range from -1 to infinity
	 * Three different cases are handled:
	 *
	 * 1. For a number of NOP cycles greater than 0, the RW Mgr looping
	 *    mechanism will be used to insert the right number of NOPs
	 *
	 * 2. For a number of NOP cycles equals to 0, the micro-instruction
	 *    issuing the write command will jump straight to the
	 *    micro-instruction that turns on DQS (for DDRx), or outputs write
	 *    data (for RLD), skipping
	 *    the NOP micro-instruction all together
	 *
	 * 3. A number of NOP cycles equal to -1 indicates that DQS must be
	 *    turned on in the same micro-instruction that issues the write
	 *    command. Then we need
	 *    to directly jump to the micro-instruction that sends out the data
	 *
	 * NOTE: Implementing this mechanism uses 2 RW Mgr jump-counters
	 *       (2 and 3). One jump-counter (0) is used to perform multiple
	 *       write-read operations.
	 *       one counter left to issue this command in "multiple-group" mode
	 */

	rw_wl_nop_cycles = gbl->rw_wl_nop_cycles;

	if (rw_wl_nop_cycles == -1) {
		/*
		 * CNTR 2 - We want to execute the special write operation that
		 * turns on DQS right away and then skip directly to the
		 * instruction that sends out the data. We set the counter to a
		 * large number so that the jump is always taken.
		 */
		writel(0xFF, &sdr_rw_load_mgr_regs->load_cntr2);

		/* CNTR 3 - Not used */
		if (test_dm) {
			mcc_instruction = RW_MGR_LFSR_WR_RD_DM_BANK_0_WL_1;
			writel(RW_MGR_LFSR_WR_RD_DM_BANK_0_DATA,
			       &sdr_rw_load_jump_mgr_regs->load_jump_add2);
			writel(RW_MGR_LFSR_WR_RD_DM_BANK_0_NOP,
			       &sdr_rw_load_jump_mgr_regs->load_jump_add3);
		} else {
			mcc_instruction = RW_MGR_LFSR_WR_RD_BANK_0_WL_1;
			writel(RW_MGR_LFSR_WR_RD_BANK_0_DATA,
				&sdr_rw_load_jump_mgr_regs->load_jump_add2);
			writel(RW_MGR_LFSR_WR_RD_BANK_0_NOP,
				&sdr_rw_load_jump_mgr_regs->load_jump_add3);
		}
	} else if (rw_wl_nop_cycles == 0) {
		/*
		 * CNTR 2 - We want to skip the NOP operation and go straight
		 * to the DQS enable instruction. We set the counter to a large
		 * number so that the jump is always taken.
		 */
		writel(0xFF, &sdr_rw_load_mgr_regs->load_cntr2);

		/* CNTR 3 - Not used */
		if (test_dm) {
			mcc_instruction = RW_MGR_LFSR_WR_RD_DM_BANK_0;
			writel(RW_MGR_LFSR_WR_RD_DM_BANK_0_DQS,
			       &sdr_rw_load_jump_mgr_regs->load_jump_add2);
		} else {
			mcc_instruction = RW_MGR_LFSR_WR_RD_BANK_0;
			writel(RW_MGR_LFSR_WR_RD_BANK_0_DQS,
				&sdr_rw_load_jump_mgr_regs->load_jump_add2);
		}
	} else {
		/*
		 * CNTR 2 - In this case we want to execute the next instruction
		 * and NOT take the jump. So we set the counter to 0. The jump
		 * address doesn't count.
		 */
		writel(0x0, &sdr_rw_load_mgr_regs->load_cntr2);
		writel(0x0, &sdr_rw_load_jump_mgr_regs->load_jump_add2);

		/*
		 * CNTR 3 - Set the nop counter to the number of cycles we
		 * need to loop for, minus 1.
		 */
		writel(rw_wl_nop_cycles - 1, &sdr_rw_load_mgr_regs->load_cntr3);
		if (test_dm) {
			mcc_instruction = RW_MGR_LFSR_WR_RD_DM_BANK_0;
			writel(RW_MGR_LFSR_WR_RD_DM_BANK_0_NOP,
				&sdr_rw_load_jump_mgr_regs->load_jump_add3);
		} else {
			mcc_instruction = RW_MGR_LFSR_WR_RD_BANK_0;
			writel(RW_MGR_LFSR_WR_RD_BANK_0_NOP,
				&sdr_rw_load_jump_mgr_regs->load_jump_add3);
		}
	}

	writel(0, SDR_PHYGRP_RWMGRGRP_ADDRESS |
		  RW_MGR_RESET_READ_DATAPATH_OFFSET);

	if (quick_write_mode)
		writel(0x08, &sdr_rw_load_mgr_regs->load_cntr0);
	else
		writel(0x40, &sdr_rw_load_mgr_regs->load_cntr0);

	writel(mcc_instruction, &sdr_rw_load_jump_mgr_regs->load_jump_add0);

	/*
	 * CNTR 1 - This is used to ensure enough time elapses
	 * for read data to come back.
	 */
	writel(0x30, &sdr_rw_load_mgr_regs->load_cntr1);

	if (test_dm) {
		writel(RW_MGR_LFSR_WR_RD_DM_BANK_0_WAIT,
			&sdr_rw_load_jump_mgr_regs->load_jump_add1);
	} else {
		writel(RW_MGR_LFSR_WR_RD_BANK_0_WAIT,
			&sdr_rw_load_jump_mgr_regs->load_jump_add1);
	}

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

/* Test writes, can check for a single bit pass or multiple bit pass */
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)
{
	uint32_t r;
	uint32_t correct_mask_vg;
	uint32_t tmp_bit_chk;
	uint32_t vg;
	uint32_t rank_end = all_ranks ? RW_MGR_MEM_NUMBER_OF_RANKS :
		(rank_bgn + NUM_RANKS_PER_SHADOW_REG);
	uint32_t addr_rw_mgr;
	uint32_t base_rw_mgr;

	*bit_chk = param->write_correct_mask;
	correct_mask_vg = param->write_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);

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

			tmp_bit_chk = tmp_bit_chk <<
				(RW_MGR_MEM_DQ_PER_WRITE_DQS /
				RW_MGR_MEM_VIRTUAL_GROUPS_PER_WRITE_DQS);
			rw_mgr_mem_calibrate_write_test_issue(write_group *
				RW_MGR_MEM_VIRTUAL_GROUPS_PER_WRITE_DQS+vg,
				use_dm);

			base_rw_mgr = readl(addr_rw_mgr);
			tmp_bit_chk = tmp_bit_chk | (correct_mask_vg & ~(base_rw_mgr));
			if (vg == 0)
				break;
		}
		*bit_chk &= tmp_bit_chk;
	}

	if (all_correct) {
		set_rank_and_odt_mask(0, RW_MGR_ODT_MODE_OFF);
		debug_cond(DLEVEL == 2, "write_test(%u,%u,ALL) : %u == \
			   %u => %lu", write_group, use_dm,
			   *bit_chk, param->write_correct_mask,
			   (long unsigned int)(*bit_chk ==
			   param->write_correct_mask));
		return *bit_chk == param->write_correct_mask;
	} else {
		set_rank_and_odt_mask(0, RW_MGR_ODT_MODE_OFF);
		debug_cond(DLEVEL == 2, "write_test(%u,%u,ONE) : %u != ",
		       write_group, use_dm, *bit_chk);
		debug_cond(DLEVEL == 2, "%lu" " => %lu", (long unsigned int)0,
			(long unsigned int)(*bit_chk != 0));
		return *bit_chk != 0x00;
	}
}

/*
 * center all windows. do per-bit-deskew to possibly increase size of
 * certain windows.
 */
static uint32_t rw_mgr_mem_calibrate_writes_center(uint32_t rank_bgn,
	uint32_t write_group, uint32_t test_bgn)
{
	uint32_t i, p, min_index;
	int32_t d;
	/*
	 * 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_WRITE_DQS];
	int32_t right_edge[RW_MGR_MEM_DQ_PER_WRITE_DQS];
	int32_t mid;
	int32_t mid_min, orig_mid_min;
	int32_t new_dqs, start_dqs, shift_dq;
	int32_t dq_margin, dqs_margin, dm_margin;
	uint32_t temp_dq_out1_delay;
	uint32_t addr;

	int ret;

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

	dm_margin = 0;

	addr = SDR_PHYGRP_SCCGRP_ADDRESS | SCC_MGR_IO_OUT1_DELAY_OFFSET;
	start_dqs = readl(addr +
			  (RW_MGR_MEM_DQ_PER_WRITE_DQS << 2));

	/* per-bit deskew */

	/*
	 * set the left and right edge of each bit to an illegal value
	 * use (IO_IO_OUT1_DELAY_MAX + 1) as an illegal value.
	 */
	sticky_bit_chk = 0;
	for (i = 0; i < RW_MGR_MEM_DQ_PER_WRITE_DQS; i++) {
		left_edge[i]  = IO_IO_OUT1_DELAY_MAX + 1;
		right_edge[i] = IO_IO_OUT1_DELAY_MAX + 1;
	}

	/* Search for the left edge of the window for each bit */
	search_left_edge(1, rank_bgn, write_group, 0, test_bgn,
			 &bit_chk, &sticky_bit_chk,
			 left_edge, right_edge, 0);

	/* Search for the right edge of the window for each bit */
	ret = search_right_edge(1, rank_bgn, write_group, 0,
				start_dqs, 0,
				&bit_chk, &sticky_bit_chk,
				left_edge, right_edge, 0);
	if (ret) {
		set_failing_group_stage(test_bgn + ret - 1, CAL_STAGE_WRITES,
					CAL_SUBSTAGE_WRITES_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_WRITE_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 write_center: mid_min=%d\n", __func__,
		   __LINE__, mid_min);

	/* Determine the amount we can change DQS (which is -mid_min) */
	orig_mid_min = mid_min;
	new_dqs = start_dqs;
	mid_min = 0;
	debug_cond(DLEVEL == 1, "%s:%d write_center: start_dqs=%d new_dqs=%d \
		   mid_min=%d\n", __func__, __LINE__, start_dqs, new_dqs, mid_min);
	/* Initialize data for export structures */
	dqs_margin = IO_IO_OUT1_DELAY_MAX + 1;
	dq_margin  = IO_IO_OUT1_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_WRITE_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, "%s:%d write_center: before: shift_dq \
			   [%u]=%d\n", __func__, __LINE__, i, shift_dq);

		addr = SDR_PHYGRP_SCCGRP_ADDRESS | SCC_MGR_IO_OUT1_DELAY_OFFSET;
		temp_dq_out1_delay = readl(addr + (i << 2));
		if (shift_dq + (int32_t)temp_dq_out1_delay >
			(int32_t)IO_IO_OUT1_DELAY_MAX) {
			shift_dq = (int32_t)IO_IO_OUT1_DELAY_MAX - temp_dq_out1_delay;
		} else if (shift_dq + (int32_t)temp_dq_out1_delay < 0) {
			shift_dq = -(int32_t)temp_dq_out1_delay;
		}
		debug_cond(DLEVEL == 2, "write_center: after: shift_dq[%u]=%d\n",
			   i, shift_dq);
		scc_mgr_set_dq_out1_delay(i, temp_dq_out1_delay + shift_dq);
		scc_mgr_load_dq(i);

		debug_cond(DLEVEL == 2, "write_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);
	}

	/* Move DQS */
	scc_mgr_apply_group_dqs_io_and_oct_out1(write_group, new_dqs);
	writel(0, &sdr_scc_mgr->update);

	/* Centre DM */
	debug_cond(DLEVEL == 2, "%s:%d write_center: DM\n", __func__, __LINE__);

	/*
	 * set the left and right edge of each bit to an illegal value,
	 * use (IO_IO_OUT1_DELAY_MAX + 1) as an illegal value,
	 */
	left_edge[0]  = IO_IO_OUT1_DELAY_MAX + 1;
	right_edge[0] = IO_IO_OUT1_DELAY_MAX + 1;
	int32_t bgn_curr = IO_IO_OUT1_DELAY_MAX + 1;
	int32_t end_curr = IO_IO_OUT1_DELAY_MAX + 1;
	int32_t bgn_best = IO_IO_OUT1_DELAY_MAX + 1;
	int32_t end_best = IO_IO_OUT1_DELAY_MAX + 1;
	int32_t win_best = 0;

	/* Search for the/part of the window with DM shift */
	for (d = IO_IO_OUT1_DELAY_MAX; d >= 0; d -= DELTA_D) {
		scc_mgr_apply_group_dm_out1_delay(d);
		writel(0, &sdr_scc_mgr->update);

		if (rw_mgr_mem_calibrate_write_test(rank_bgn, write_group, 1,
						    PASS_ALL_BITS, &bit_chk,
						    0)) {
			/* USE Set current end of the window */
			end_curr = -d;
			/*
			 * If a starting edge of our window has not been seen
			 * this is our current start of the DM window.
			 */
			if (bgn_curr == IO_IO_OUT1_DELAY_MAX + 1)
				bgn_curr = -d;

			/*
			 * If current window is bigger than best seen.
			 * Set best seen to be current window.
			 */
			if ((end_curr-bgn_curr+1) > win_best) {
				win_best = end_curr-bgn_curr+1;
				bgn_best = bgn_curr;
				end_best = end_curr;
			}
		} else {
			/* We just saw a failing test. Reset temp edge */
			bgn_curr = IO_IO_OUT1_DELAY_MAX + 1;
			end_curr = IO_IO_OUT1_DELAY_MAX + 1;