path: root/config/coreboot/default/patches/0056-haswell-NRI-Add-write-leveling.patch

From ce0ed94f993506e75b711c214b49ba480037e7d3 Mon Sep 17 00:00:00 2001
From: Angel Pons <th3fanbus@gmail.com>
Date: Sun, 8 May 2022 12:56:04 +0200
Subject: [PATCH 14/17] haswell NRI: Add write leveling

Implement JEDEC write leveling, which is done in two steps. The first
step uses the JEDEC procedure to do "fine" write leveling, i.e. align
the DQS phase to the clock signal. The second step performs a regular
read-write test to correct "coarse" cycle errors.

Change-Id: I27678523fe22c38173a688e2a4751c259a20f009
Signed-off-by: Angel Pons <th3fanbus@gmail.com>
---
 .../intel/haswell/native_raminit/Makefile.mk  |   1 +
 .../haswell/native_raminit/raminit_main.c     |   1 +
 .../haswell/native_raminit/raminit_native.h   |  10 +
 .../train_jedec_write_leveling.c              | 581 ++++++++++++++++++
 .../intel/haswell/registers/mchbar.h          |   2 +
 5 files changed, 595 insertions(+)
 create mode 100644 src/northbridge/intel/haswell/native_raminit/train_jedec_write_leveling.c

diff --git a/src/northbridge/intel/haswell/native_raminit/Makefile.mk b/src/northbridge/intel/haswell/native_raminit/Makefile.mk
index c442be0728..40c2f5e014 100644
--- a/src/northbridge/intel/haswell/native_raminit/Makefile.mk
+++ b/src/northbridge/intel/haswell/native_raminit/Makefile.mk
@@ -16,5 +16,6 @@ romstage-y += setup_wdb.c
 romstage-y += spd_bitmunching.c
 romstage-y += testing_io.c
 romstage-y += timings_refresh.c
+romstage-y += train_jedec_write_leveling.c
 romstage-y += train_read_mpr.c
 romstage-y += train_receive_enable.c
diff --git a/src/northbridge/intel/haswell/native_raminit/raminit_main.c b/src/northbridge/intel/haswell/native_raminit/raminit_main.c
index 264d1468f5..1ff23be615 100644
--- a/src/northbridge/intel/haswell/native_raminit/raminit_main.c
+++ b/src/northbridge/intel/haswell/native_raminit/raminit_main.c
@@ -62,6 +62,7 @@ static const struct task_entry cold_boot[] = {
 	{ pre_training,                                           true, "PRETRAIN",   },
 	{ train_receive_enable,                                   true, "RCVET",      },
 	{ train_read_mpr,                                         true, "RDMPRT",     },
+	{ train_jedec_write_leveling,                             true, "JWRL",       },
 };
 
 /* Return a generic stepping value to make stepping checks simpler */
diff --git a/src/northbridge/intel/haswell/native_raminit/raminit_native.h b/src/northbridge/intel/haswell/native_raminit/raminit_native.h
index 7a486479ea..d6b11b9d3c 100644
--- a/src/northbridge/intel/haswell/native_raminit/raminit_native.h
+++ b/src/northbridge/intel/haswell/native_raminit/raminit_native.h
@@ -60,6 +60,9 @@
 /* Specified in PI ticks. 64 PI ticks == 1 qclk */
 #define tDQSCK_DRIFT		64
 
+/* Maximum additional latency */
+#define MAX_ADD_DELAY		2
+
 enum margin_parameter {
 	RcvEna,
 	RdT,
@@ -216,6 +219,7 @@ enum raminit_status {
 	RAMINIT_STATUS_REUT_ERROR,
 	RAMINIT_STATUS_RCVEN_FAILURE,
 	RAMINIT_STATUS_RMPR_FAILURE,
+	RAMINIT_STATUS_JWRL_FAILURE,
 	RAMINIT_STATUS_UNSPECIFIED_ERROR, /** TODO: Deprecated in favor of specific values **/
 };
 
@@ -381,6 +385,11 @@ static inline uint32_t get_data_train_feedback(const uint8_t channel, const uint
 	return mchbar_read32(DDR_DATA_TRAIN_FEEDBACK(channel, byte));
 }
 
+static inline uint16_t get_byte_group_errors(const uint8_t channel)
+{
+	return mchbar_read32(4 + REUT_ch_ERR_MISC_STATUS(channel)) & 0x1ff;
+}
+
 /* Number of ticks to wait in units of 69.841279 ns (citation needed) */
 static inline void tick_delay(const uint32_t delay)
 {
@@ -438,6 +447,7 @@ enum raminit_status configure_memory_map(struct sysinfo *ctrl);
 enum raminit_status do_jedec_init(struct sysinfo *ctrl);
 enum raminit_status train_receive_enable(struct sysinfo *ctrl);
 enum raminit_status train_read_mpr(struct sysinfo *ctrl);
+enum raminit_status train_jedec_write_leveling(struct sysinfo *ctrl);
 
 void configure_timings(struct sysinfo *ctrl);
 void configure_refresh(struct sysinfo *ctrl);
diff --git a/src/northbridge/intel/haswell/native_raminit/train_jedec_write_leveling.c b/src/northbridge/intel/haswell/native_raminit/train_jedec_write_leveling.c
new file mode 100644
index 0000000000..ef6483e2bd
--- /dev/null
+++ b/src/northbridge/intel/haswell/native_raminit/train_jedec_write_leveling.c
@@ -0,0 +1,581 @@
+/* SPDX-License-Identifier: GPL-2.0-or-later */
+
+#include <assert.h>
+#include <console/console.h>
+#include <delay.h>
+#include <northbridge/intel/haswell/haswell.h>
+#include <types.h>
+
+#include "raminit_native.h"
+#include "ranges.h"
+
+#define JWLC_PLOT	RAM_DEBUG
+#define JWRL_PLOT	RAM_DEBUG
+
+static void reset_dram_dll(struct sysinfo *ctrl, const uint8_t channel, const uint8_t rank)
+{
+	const uint16_t mr0reg = ctrl->mr0[channel][rank / 2];
+	reut_issue_mrs(ctrl, channel, BIT(rank), 0, mr0reg | MR0_DLL_RESET);
+}
+
+static void program_wdb_pattern(struct sysinfo *ctrl, const bool invert)
+{
+	/* Pattern to keep DQ-DQS simple but detect any failures. Same as NHM/WSM. */
+	const uint8_t pat[4][2] = {
+		{ 0x00, 0xff },
+		{ 0xff, 0x00 },
+		{ 0xc3, 0x3c },
+		{ 0x3c, 0xc3 },
+	};
+	const uint8_t pmask[2][8] = {
+		{ 0, 0, 1, 1, 1, 1, 0, 0 },
+		{ 1, 1, 0, 0, 0, 0, 1, 1 },
+	};
+	for (uint8_t s = 0; s < ARRAY_SIZE(pat); s++)
+		write_wdb_fixed_pat(ctrl, pat[s], pmask[invert], ARRAY_SIZE(pmask[invert]), s);
+}
+
+static int16_t set_add_delay(uint32_t *add_delay, uint8_t rank, int8_t target_off)
+{
+	const uint8_t shift = rank * 2;
+	if (target_off > MAX_ADD_DELAY) {
+		*add_delay &= ~(3 << shift);
+		*add_delay |= MAX_ADD_DELAY << shift;
+		return 128 * (target_off - MAX_ADD_DELAY);
+	} else if (target_off < 0) {
+		*add_delay &= ~(3 << shift);
+		*add_delay |= 0 << shift;
+		return 128 * target_off;
+	} else {
+		*add_delay &= ~(3 << shift);
+		*add_delay |= target_off << shift;
+		return 0;
+	}
+}
+
+static enum raminit_status train_jedec_write_leveling_cleanup(struct sysinfo *ctrl)
+{
+	const struct reut_box reut_addr = {
+		.col = {
+			.start   = 0,
+			.stop    = 1023,
+			.inc_val = 1,
+		},
+	};
+	const struct wdb_pat wdb_pattern = {
+		.start_ptr  = 0,
+		.stop_ptr   = 3,
+		.inc_rate   = 1,
+		.dq_pattern = BASIC_VA,
+	};
+	const int8_t offsets[] = { 0, 1, -1, 2, 3 };
+	const int8_t dq_offsets[] = { 0, -10, 10, -5, 5, -15, 15 };
+	const uint8_t dq_offset_max = ARRAY_SIZE(dq_offsets);
+
+	/* Set LFSR seeds to be sequential */
+	program_wdb_lfsr(ctrl, true);
+	setup_io_test(
+		ctrl,
+		ctrl->chanmap,
+		PAT_WR_RD,
+		2,
+		4,
+		&reut_addr,
+		NSOE,
+		&wdb_pattern,
+		0,
+		0);
+
+	const union reut_pat_wdb_cl_mux_cfg_reg reut_wdb_cl_mux_cfg = {
+		.mux_0_control       = REUT_MUX_BTBUFFER,
+		.mux_1_control       = REUT_MUX_BTBUFFER,
+		.mux_2_control       = REUT_MUX_BTBUFFER,
+		.ecc_data_source_sel = 1,
+	};
+	for (uint8_t channel = 0; channel < NUM_CHANNELS; channel++) {
+		if (!does_ch_exist(ctrl, channel))
+			continue;
+
+		mchbar_write32(REUT_ch_PAT_WDB_CL_MUX_CFG(channel), reut_wdb_cl_mux_cfg.raw);
+	}
+
+	int8_t byte_off[NUM_CHANNELS][NUM_LANES] = { 0 };
+	uint32_t add_delay[NUM_CHANNELS] = { 0 };
+	enum raminit_status status = RAMINIT_STATUS_SUCCESS;
+	bool invert = false;
+	const uint16_t valid_byte_mask = BIT(ctrl->lanes) - 1;
+	for (uint8_t rank = 0; rank < NUM_SLOTRANKS; rank++) {
+		uint8_t chanmask = 0;
+		for (uint8_t channel = 0; channel < NUM_CHANNELS; channel++)
+			chanmask |= select_reut_ranks(ctrl, channel, BIT(rank));
+
+		if (!chanmask)
+			continue;
+
+		printk(BIOS_DEBUG, "Rank %u\n", rank);
+		printk(JWLC_PLOT, "Channel");
+		for (uint8_t channel = 0; channel < NUM_CHANNELS; channel++) {
+			if (!rank_in_ch(ctrl, rank, channel))
+				continue;
+
+			printk(JWLC_PLOT, "\t\t%u\t", channel);
+		}
+		printk(JWLC_PLOT, "\nByte\t");
+		for (uint8_t channel = 0; channel < NUM_CHANNELS; channel++) {
+			if (!rank_in_ch(ctrl, rank, channel))
+				continue;
+
+			printk(JWLC_PLOT, "\t");
+			for (uint8_t byte = 0; byte < ctrl->lanes; byte++)
+				printk(JWLC_PLOT, "%u ", byte);
+		}
+		printk(JWLC_PLOT, "\nDelay DqOffset");
+		bool done = false;
+		int8_t byte_sum[NUM_CHANNELS] = { 0 };
+		uint16_t byte_pass[NUM_CHANNELS] = { 0 };
+		for (uint8_t off = 0; off < ARRAY_SIZE(offsets); off++) {
+			for (uint8_t channel = 0; channel < NUM_CHANNELS; channel++) {
+				if (!rank_in_ch(ctrl, rank, channel))
+					continue;
+
+				const int16_t global_byte_off =
+					set_add_delay(&add_delay[channel], rank, offsets[off]);
+				for (uint8_t byte = 0; byte < ctrl->lanes; byte++) {
+					update_txt(ctrl, channel, rank, byte, TXT_DQDQS_OFF,
+						global_byte_off);
+				}
+				mchbar_write32(SC_WR_ADD_DELAY_ch(channel),
+						add_delay[channel]);
+			}
+			/* Reset FIFOs and DRAM DLL (Micron workaround) */
+			if (!ctrl->lpddr) {
+				io_reset();
+				for (uint8_t channel = 0; channel < NUM_CHANNELS; channel++) {
+					if (!rank_in_ch(ctrl, rank, channel))
+						continue;
+
+					reset_dram_dll(ctrl, channel, rank);
+				}
+				udelay(1);
+			}
+			for (uint8_t dq_offset = 0; dq_offset < dq_offset_max; dq_offset++) {
+				printk(JWLC_PLOT, "\n% 3d\t% 3d",
+					offsets[off], dq_offsets[dq_offset]);
+				change_1d_margin_multicast(
+					ctrl,
+					WrT,
+					dq_offsets[dq_offset],
+					rank,
+					false,
+					REG_FILE_USE_RANK);
+
+				/*
+				 * Re-program the WDB pattern. Change the pattern
+				 * for the next test to avoid false pass issues.
+				 */
+				program_wdb_pattern(ctrl, invert);
+				invert = !invert;
+				run_io_test(ctrl, chanmask, BASIC_VA, true);
+				done = true;
+				for (uint8_t channel = 0; channel < NUM_CHANNELS; channel++) {
+					if (!rank_in_ch(ctrl, rank, channel))
+						continue;
+
+					printk(JWLC_PLOT, "\t");
+					uint16_t result = get_byte_group_errors(channel);
+					result &= valid_byte_mask;
+
+					/* Skip bytes that have failed or already passed */
+					const uint16_t skip_me = result | byte_pass[channel];
+					for (uint8_t byte = 0; byte < ctrl->lanes; byte++) {
+						const bool pass = result & BIT(byte);
+						printk(JWLC_PLOT, pass ? "# " : ". ");
+						if (skip_me & BIT(byte))
+							continue;
+
+						byte_pass[channel] |= BIT(byte);
+						byte_off[channel][byte] = offsets[off];
+						byte_sum[channel] += offsets[off];
+					}
+					if (byte_pass[channel] != valid_byte_mask)
+						done = false;
+				}
+				if (done)
+					break;
+			}
+			if (done)
+				break;
+		}
+		printk(BIOS_DEBUG, "\n\n");
+		if (!done) {
+			printk(BIOS_ERR, "JWLC: Could not find a pass for all bytes\n");
+			for (uint8_t channel = 0; channel < NUM_CHANNELS; channel++) {
+				if (!rank_in_ch(ctrl, rank, channel))
+					continue;
+
+				printk(BIOS_ERR, "Channel %u, rank %u fail:", channel, rank);
+				const uint16_t passing_mask = byte_pass[channel];
+				for (uint8_t byte = 0; byte < ctrl->lanes; byte++) {
+					if (BIT(byte) & passing_mask)
+						continue;
+
+					printk(BIOS_ERR, " %u", byte);
+				}
+				printk(BIOS_ERR, "\n");
+			}
+			status = RAMINIT_STATUS_JWRL_FAILURE;
+			break;
+		}
+		for (uint8_t channel = 0; channel < NUM_CHANNELS; channel++) {
+			if (!rank_in_ch(ctrl, rank, channel))
+				continue;
+
+			/* Refine target offset to make sure it works for all bytes */
+			int8_t target_off = DIV_ROUND_CLOSEST(byte_sum[channel], ctrl->lanes);
+			int16_t global_byte_off = 0;
+			uint8_t all_good_loops = 0;
+			bool all_good = 0;
+			while (!all_good) {
+				global_byte_off =
+					set_add_delay(&add_delay[channel], rank, target_off);
+				all_good = true;
+				for (uint8_t byte = 0; byte < ctrl->lanes; byte++) {
+					int16_t local_offset;
+					local_offset = byte_off[channel][byte] - target_off;
+					local_offset = local_offset * 128 + global_byte_off;
+					const uint16_t tx_dq = ctrl->tx_dq[channel][rank][byte];
+					if (tx_dq + local_offset >= (512 - 64)) {
+						all_good = false;
+						all_good_loops++;
+						target_off++;
+						break;
+					}
+					const uint16_t txdqs = ctrl->tx_dq[channel][rank][byte];
+					if (txdqs + local_offset < 96) {
+						all_good = false;
+						all_good_loops++;
+						target_off--;
+						break;
+					}
+				}
+				/* Avoid an infinite loop */
+				if (all_good_loops > 3)
+					break;
+			}
+			if (!all_good) {
+				printk(BIOS_ERR, "JWLC: Target offset refining failed\n");
+				status = RAMINIT_STATUS_JWRL_FAILURE;
+				break;
+			}
+			printk(BIOS_DEBUG, "C%u.R%u:  Offset\tFinalEdge\n", channel, rank);
+			for (uint8_t byte = 0; byte < ctrl->lanes; byte++) {
+				int16_t local_offset;
+				local_offset = byte_off[channel][byte] - target_off;
+				local_offset = local_offset * 128 + global_byte_off;
+				ctrl->tx_dq[channel][rank][byte] += local_offset;
+				ctrl->txdqs[channel][rank][byte] += local_offset;
+				update_txt(ctrl, channel, rank, byte, TXT_RESTORE, 0);
+				printk(BIOS_DEBUG, "  B%u:   %d\t%d\n", byte, local_offset,
+					ctrl->txdqs[channel][rank][byte]);
+			}
+			mchbar_write32(SC_WR_ADD_DELAY_ch(channel), add_delay[channel]);
+			if (!ctrl->lpddr) {
+				reset_dram_dll(ctrl, channel, rank);
+				udelay(1);
+			}
+			printk(BIOS_DEBUG, "\n");
+		}
+		printk(BIOS_DEBUG, "\n");
+	}
+
+	/* Restore WDB after test */
+	write_wdb_va_pat(ctrl, 0, BASIC_VA_PAT_SPREAD_8, 8, 0);
+	program_wdb_lfsr(ctrl, false);
+	mchbar_write32(DDR_DATA_OFFSET_TRAIN, 0);
+
+	/** TODO: Do full JEDEC init instead? **/
+	io_reset();
+	return status;
+}
+
+static enum raminit_status verify_wl_width(const int32_t lwidth)
+{
+	if (lwidth <= 32) {
+		/* Check if width is valid */
+		printk(BIOS_ERR, "WrLevel: Width region (%d) too small\n", lwidth);
+		return RAMINIT_STATUS_JWRL_FAILURE;
+	}
+	if (lwidth >= 96) {
+		/* Since we're calibrating a phase, a too large region is a problem */
+		printk(BIOS_ERR, "WrLevel: Width region (%d) too large\n", lwidth);
+		return RAMINIT_STATUS_JWRL_FAILURE;
+	}
+	return 0;
+}
+
+enum raminit_status train_jedec_write_leveling(struct sysinfo *ctrl)
+{
+	/*
+	 * Enabling WL mode causes DQS to toggle for 1024 QCLK.
+	 * Wait for this to stop. Round up to nearest microsecond.
+	 */
+	const bool wl_long_delay = ctrl->lpddr;
+	const uint32_t dqs_toggle_time = wl_long_delay ? 2048 : 1024;
+	const uint32_t wait_time_us = DIV_ROUND_UP(ctrl->qclkps * dqs_toggle_time, 1000 * 1000);
+
+	const uint16_t wl_start = 192;
+	const uint16_t wl_stop  = 192 + 128;
+	const uint16_t wl_step  = 2;
+
+	/* Do not use cached MR values */
+	const bool save_restore_mrs = ctrl->restore_mrs;
+	ctrl->restore_mrs = 0;
+
+	/* Propagate delay values (without a write command) */
+	for (uint8_t channel = 0; channel < NUM_CHANNELS; channel++) {
+		if (!does_ch_exist(ctrl, channel))
+			continue;
+
+		/* Propagate delay values from rank 0 to prevent assertion failures in RTL */
+		union ddr_data_control_0_reg data_control_0 = {
+			.raw = ctrl->dq_control_0[channel],
+		};
+		data_control_0.read_rf_rd   = 0;
+		data_control_0.read_rf_wr   = 1;
+		data_control_0.read_rf_rank = 0;
+		mchbar_write32(DDR_DATA_ch_CONTROL_0(channel), data_control_0.raw);
+		for (uint8_t byte = 0; byte < ctrl->lanes; byte++) {
+			union ddr_data_control_2_reg data_control_2 = {
+				.raw = ctrl->dq_control_2[channel][byte],
+			};
+			data_control_2.force_bias_on = 1;
+			data_control_2.force_rx_on   = 0;
+			data_control_2.wl_long_delay = wl_long_delay;
+			mchbar_write32(DQ_CONTROL_2(channel, byte), data_control_2.raw);
+		}
+	}
+
+	if (ctrl->lpddr)
+		die("%s: Missing LPDDR support\n", __func__);
+
+	if (!ctrl->lpddr)
+		ddr3_program_mr1(ctrl, 0, 1);
+
+	enum raminit_status status = RAMINIT_STATUS_SUCCESS;
+	struct phase_train_data region_data[NUM_CHANNELS][NUM_LANES] = { 0 };
+	for (uint8_t rank = 0; rank < NUM_SLOTRANKS; rank++) {
+		if (!does_rank_exist(ctrl, rank))
+			continue;
+
+		for (uint8_t channel = 0; channel < NUM_CHANNELS; channel++) {
+			if (!rank_in_ch(ctrl, rank, channel))
+				continue;
+
+			/** TODO: Differs for LPDDR **/
+			uint16_t mr1reg = ctrl->mr1[channel][rank / 2];
+			mr1reg &= ~MR1_QOFF_ENABLE;
+			mr1reg |= MR1_WL_ENABLE;
+			if (is_hsw_ult()) {
+				mr1reg &= ~RTTNOM_MASK;
+				mr1reg |= encode_ddr3_rttnom(120);
+			} else if (ctrl->dpc[channel] == 2) {
+				mr1reg &= ~RTTNOM_MASK;
+				mr1reg |= encode_ddr3_rttnom(60);
+			}
+			reut_issue_mrs(ctrl, channel, BIT(rank), 1, mr1reg);
+
+			/* Assert ODT for myself */
+			uint8_t odt_matrix = BIT(rank);
+			if (ctrl->dpc[channel] == 2) {
+				/* Assert ODT for non-target DIMM */
+				const uint8_t other_dimm = ((rank + 2) / 2) & 1;
+				odt_matrix |= BIT(2 * other_dimm);
+			}
+
+			union reut_misc_odt_ctrl_reg reut_misc_odt_ctrl = {
+				.raw = 0,
+			};
+			if (ctrl->lpddr) {
+				/* Only one ODT pin for ULT */
+				reut_misc_odt_ctrl.odt_on       = 1;
+				reut_misc_odt_ctrl.odt_override = 1;
+			} else if (!is_hsw_ult()) {
+				reut_misc_odt_ctrl.odt_on       = odt_matrix;
+				reut_misc_odt_ctrl.odt_override = 0xf;
+			}
+			mchbar_write32(REUT_ch_MISC_ODT_CTRL(channel), reut_misc_odt_ctrl.raw);
+		}
+		for (uint8_t channel = 0; channel < NUM_CHANNELS; channel++) {
+			if (!rank_in_ch(ctrl, rank, channel))
+				continue;
+
+			/*
+			 * Enable write leveling mode in DDR and propagate delay
+			 * values (without a write command). Stay in WL mode.
+			 */
+			union ddr_data_control_0_reg data_control_0 = {
+				.raw = ctrl->dq_control_0[channel],
+			};
+			data_control_0.wl_training_mode = 1;
+			data_control_0.tx_pi_on         = 1;
+			data_control_0.read_rf_rd       = 0;
+			data_control_0.read_rf_wr       = 1;
+			data_control_0.read_rf_rank     = rank;
+			mchbar_write32(DDR_DATA_ch_CONTROL_0(channel), data_control_0.raw);
+		}
+		printk(BIOS_DEBUG, "\nRank %u\n", rank);
+		printk(JWRL_PLOT, "Channel\t");
+		for (uint8_t channel = 0; channel < NUM_CHANNELS; channel++) {
+			if (!rank_in_ch(ctrl, rank, channel))
+				continue;
+
+			printk(JWRL_PLOT, "%u", channel);
+			if (channel > 0)
+				continue;
+
+			for (uint8_t byte = 0; byte < ctrl->lanes; byte++)
+				printk(JWRL_PLOT, "\t");
+		}
+		printk(JWRL_PLOT, "\nByte");
+		for (uint8_t channel = 0; channel < NUM_CHANNELS; channel++) {
+			if (!rank_in_ch(ctrl, rank, channel))
+				continue;
+
+			for (uint8_t byte = 0; byte < ctrl->lanes; byte++)
+				printk(JWRL_PLOT, "\t%u", byte);
+		}
+		printk(JWRL_PLOT, "\nWlDelay");
+		for (uint16_t wl_delay = wl_start; wl_delay < wl_stop; wl_delay += wl_step) {
+			printk(JWRL_PLOT, "\n %3u:", wl_delay);
+			for (uint8_t channel = 0; channel < NUM_CHANNELS; channel++) {
+				if (!rank_in_ch(ctrl, rank, channel))
+					continue;
+
+				for (uint8_t byte = 0; byte < ctrl->lanes; byte++) {
+					update_txt(ctrl, channel, rank, byte, TXT_TXDQS,
+						wl_delay);
+				}
+			}
+			/* Wait for the first burst to finish */
+			if (wl_delay == wl_start)
+				udelay(wait_time_us);
+
+			io_reset();
+			udelay(wait_time_us);
+			for (uint8_t channel = 0; channel < NUM_CHANNELS; channel++) {
+				if (!rank_in_ch(ctrl, rank, channel))
+					continue;
+
+				for (uint8_t byte = 0; byte < ctrl->lanes; byte++) {
+					const uint32_t feedback =
+						get_data_train_feedback(channel, byte);
+					const bool pass = (feedback & 0x1ff) >= 16;
+					printk(JWRL_PLOT, "\t%c%u", pass ? '.' : '#', feedback);
+					phase_record_pass(
+						&region_data[channel][byte],
+						pass,
+						wl_delay,
+						wl_start,
+						wl_step);
+				}
+			}
+		}
+		printk(JWRL_PLOT, "\n");
+		printk(BIOS_DEBUG, "\n\tInitSt\tInitEn\tCurrSt\tCurrEn\tLargSt\tLargEn\n");
+		for (uint8_t channel = 0; channel < NUM_CHANNELS; channel++) {
+			if (!rank_in_ch(ctrl, rank, channel))
+				continue;
+
+			printk(BIOS_DEBUG, "C%u\n", channel);
+			for (uint8_t byte = 0; byte < ctrl->lanes; byte++) {
+				struct phase_train_data *data = &region_data[channel][byte];
+
+				phase_append_initial_to_current(data, wl_start, wl_step);
+				printk(BIOS_DEBUG, "   B%u:\t%d\t%d\t%d\t%d\t%d\t%d\n",
+					byte,
+					data->initial.start,
+					data->initial.end,
+					data->current.start,
+					data->current.end,
+					data->largest.start,
+					data->largest.end);
+			}
+		}
+
+		/*
+		 * Clean up after test. Very coarsely adjust for
+		 * any cycle errors. Program values for TxDQS.
+		 */
+		for (uint8_t channel = 0; channel < NUM_CHANNELS; channel++) {
+			if (!rank_in_ch(ctrl, rank, channel))
+				continue;
+
+			/* Clear ODT before MRS (JEDEC spec) */
+			mchbar_write32(REUT_ch_MISC_ODT_CTRL(channel), 0);
+
+			/** TODO: Differs for LPDDR **/
+			const uint16_t mr1reg = ctrl->mr1[channel][rank / 2] | MR1_QOFF_ENABLE;
+			reut_issue_mrs(ctrl, channel, BIT(rank), 1, mr1reg);
+
+			printk(BIOS_DEBUG, "\nC%u.R%u:  LftEdge Width\n", channel, rank);
+			const bool rank_x16 = ctrl->dimms[channel][rank / 2].data.width == 16;
+			for (uint8_t byte = 0; byte < ctrl->lanes; byte++) {
+				struct phase_train_data *data = &region_data[channel][byte];
+				const int32_t lwidth = range_width(data->largest);
+				int32_t tx_start = data->largest.start;
+				printk(BIOS_DEBUG, "  B%u:   %d\t%d\n", byte, tx_start, lwidth);
+				status = verify_wl_width(lwidth);
+				if (status) {
+					printk(BIOS_ERR,
+						"WrLevel problems on channel %u, byte %u\n",
+						channel, byte);
+					goto clean_up;
+				}
+
+				/* Align byte pairs if DIMM is x16 */
+				if (rank_x16 && (byte & 1)) {
+					const struct phase_train_data *const ref_data =
+							&region_data[channel][byte - 1];
+
+					if (tx_start > ref_data->largest.start + 64)
+						tx_start -= 128;
+
+					if (tx_start < ref_data->largest.start - 64)
+						tx_start += 128;
+				}
+
+				/* Fix for b4618067 - need to add 1 QCLK to DQS PI */
+				if (is_hsw_ult())
+					tx_start += 64;
+
+				assert(tx_start >= 0);
+				ctrl->txdqs[channel][rank][byte] = tx_start;
+				ctrl->tx_dq[channel][rank][byte] = tx_start + 32;
+				update_txt(ctrl, channel, rank, byte, TXT_RESTORE, 0);
+			}
+		}
+		printk(BIOS_DEBUG, "\n");
+	}
+
+clean_up:
+	for (uint8_t channel = 0; channel < NUM_CHANNELS; channel++) {
+		if (!does_ch_exist(ctrl, channel))
+			continue;
+
+		mchbar_write32(DDR_DATA_ch_CONTROL_0(channel), ctrl->dq_control_0[channel]);
+		for (uint8_t byte = 0; byte < ctrl->lanes; byte++) {
+			mchbar_write32(DQ_CONTROL_2(channel, byte),
+				ctrl->dq_control_2[channel][byte]);
+		}
+	}
+	if (!ctrl->lpddr)
+		ddr3_program_mr1(ctrl, 0, 0);
+
+	ctrl->restore_mrs = save_restore_mrs;
+
+	if (status)
+		return status;
+
+	/** TODO: If this step fails and dec_wrd is set, clear it and try again **/
+	return train_jedec_write_leveling_cleanup(ctrl);
+}
diff --git a/src/northbridge/intel/haswell/registers/mchbar.h b/src/northbridge/intel/haswell/registers/mchbar.h
index 6a31d3a32c..7c0b5a49de 100644
--- a/src/northbridge/intel/haswell/registers/mchbar.h
+++ b/src/northbridge/intel/haswell/registers/mchbar.h
@@ -121,6 +121,8 @@
 
 #define REUT_ch_ERR_DATA_MASK(ch)		_MCMAIN_C(0x40d8, ch)
 
+#define REUT_ch_ERR_MISC_STATUS(ch)		_MCMAIN_C(0x40e8, ch)
+
 #define REUT_ch_MISC_CKE_CTRL(ch)		_MCMAIN_C(0x4190, ch)
 #define REUT_ch_MISC_ODT_CTRL(ch)		_MCMAIN_C(0x4194, ch)
 #define REUT_ch_MISC_PAT_CADB_CTRL(ch)		_MCMAIN_C(0x4198, ch)
-- 
2.39.2