summaryrefslogtreecommitdiff
path: root/config/coreboot/default/patches/0053-haswell-NRI-Add-RcvEn-training.patch
diff options
context:
space:
mode:
Diffstat (limited to 'config/coreboot/default/patches/0053-haswell-NRI-Add-RcvEn-training.patch')
-rw-r--r--config/coreboot/default/patches/0053-haswell-NRI-Add-RcvEn-training.patch708
1 files changed, 708 insertions, 0 deletions
diff --git a/config/coreboot/default/patches/0053-haswell-NRI-Add-RcvEn-training.patch b/config/coreboot/default/patches/0053-haswell-NRI-Add-RcvEn-training.patch
new file mode 100644
index 00000000..a9821796
--- /dev/null
+++ b/config/coreboot/default/patches/0053-haswell-NRI-Add-RcvEn-training.patch
@@ -0,0 +1,708 @@
+From 0826d1e9ba50daad13c3d5adccba4b180c82296b Mon Sep 17 00:00:00 2001
+From: Angel Pons <th3fanbus@gmail.com>
+Date: Sun, 8 May 2022 00:05:41 +0200
+Subject: [PATCH 11/17] haswell NRI: Add RcvEn training
+
+Implement the RcvEn (Receive Enable) calibration procedure.
+
+Change-Id: Ifbfa520f3e0486c56d0988ce67af2ddb9cf29888
+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 | 14 +
+ .../haswell/native_raminit/reg_structs.h | 13 +
+ .../native_raminit/train_receive_enable.c | 561 ++++++++++++++++++
+ .../intel/haswell/registers/mchbar.h | 3 +
+ 6 files changed, 593 insertions(+)
+ create mode 100644 src/northbridge/intel/haswell/native_raminit/train_receive_enable.c
+
+diff --git a/src/northbridge/intel/haswell/native_raminit/Makefile.mk b/src/northbridge/intel/haswell/native_raminit/Makefile.mk
+index ebe9e9b762..e2fbfb4211 100644
+--- a/src/northbridge/intel/haswell/native_raminit/Makefile.mk
++++ b/src/northbridge/intel/haswell/native_raminit/Makefile.mk
+@@ -16,3 +16,4 @@ romstage-y += setup_wdb.c
+ romstage-y += spd_bitmunching.c
+ romstage-y += testing_io.c
+ romstage-y += timings_refresh.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 5e4674957d..7d444659c3 100644
+--- a/src/northbridge/intel/haswell/native_raminit/raminit_main.c
++++ b/src/northbridge/intel/haswell/native_raminit/raminit_main.c
+@@ -60,6 +60,7 @@ static const struct task_entry cold_boot[] = {
+ { configure_memory_map, true, "MEMMAP", },
+ { do_jedec_init, true, "JEDECINIT", },
+ { pre_training, true, "PRETRAIN", },
++ { train_receive_enable, true, "RCVET", },
+ };
+
+ /* 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 8707257b27..eaaaedad1e 100644
+--- a/src/northbridge/intel/haswell/native_raminit/raminit_native.h
++++ b/src/northbridge/intel/haswell/native_raminit/raminit_native.h
+@@ -43,6 +43,9 @@
+ #define NUM_WDB_CL_MUX_SEEDS 3
+ #define NUM_CADB_MUX_SEEDS 3
+
++/* Specified in PI ticks. 64 PI ticks == 1 qclk */
++#define tDQSCK_DRIFT 64
++
+ /* ZQ calibration types */
+ enum {
+ ZQ_INIT, /* DDR3: ZQCL with tZQinit, LPDDR3: ZQ Init with tZQinit */
+@@ -189,6 +192,7 @@ enum raminit_status {
+ RAMINIT_STATUS_MPLL_INIT_FAILURE,
+ RAMINIT_STATUS_POLL_TIMEOUT,
+ RAMINIT_STATUS_REUT_ERROR,
++ RAMINIT_STATUS_RCVEN_FAILURE,
+ RAMINIT_STATUS_UNSPECIFIED_ERROR, /** TODO: Deprecated in favor of specific values **/
+ };
+
+@@ -271,6 +275,10 @@ struct sysinfo {
+
+ union ddr_data_vref_adjust_reg dimm_vref;
+
++ uint8_t io_latency[NUM_CHANNELS][NUM_SLOTRANKS];
++ uint8_t rt_latency[NUM_CHANNELS][NUM_SLOTRANKS];
++ uint32_t rt_io_comp[NUM_CHANNELS];
++
+ uint32_t data_offset_train[NUM_CHANNELS][NUM_LANES];
+ uint32_t data_offset_comp[NUM_CHANNELS][NUM_LANES];
+
+@@ -345,6 +353,11 @@ static inline void clear_data_offset_train_all(struct sysinfo *ctrl)
+ memset(ctrl->data_offset_train, 0, sizeof(ctrl->data_offset_train));
+ }
+
++static inline uint32_t get_data_train_feedback(const uint8_t channel, const uint8_t byte)
++{
++ return mchbar_read32(DDR_DATA_TRAIN_FEEDBACK(channel, byte));
++}
++
+ /* Number of ticks to wait in units of 69.841279 ns (citation needed) */
+ static inline void tick_delay(const uint32_t delay)
+ {
+@@ -400,6 +413,7 @@ enum raminit_status convert_timings(struct sysinfo *ctrl);
+ enum raminit_status configure_mc(struct sysinfo *ctrl);
+ 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);
+
+ void configure_timings(struct sysinfo *ctrl);
+ void configure_refresh(struct sysinfo *ctrl);
+diff --git a/src/northbridge/intel/haswell/native_raminit/reg_structs.h b/src/northbridge/intel/haswell/native_raminit/reg_structs.h
+index b943259b91..b099f4bb82 100644
+--- a/src/northbridge/intel/haswell/native_raminit/reg_structs.h
++++ b/src/northbridge/intel/haswell/native_raminit/reg_structs.h
+@@ -297,6 +297,19 @@ union ddr_scram_misc_control_reg {
+ uint32_t raw;
+ };
+
++union sc_io_latency_reg {
++ struct __packed {
++ uint32_t iolat_rank0 : 4; // Bits 3:0
++ uint32_t iolat_rank1 : 4; // Bits 7:4
++ uint32_t iolat_rank2 : 4; // Bits 11:8
++ uint32_t iolat_rank3 : 4; // Bits 15:12
++ uint32_t rt_iocomp : 6; // Bits 21:16
++ uint32_t : 9; // Bits 30:22
++ uint32_t dis_rt_clk_gate : 1; // Bits 31:31
++ };
++ uint32_t raw;
++};
++
+ union mcscheds_cbit_reg {
+ struct __packed {
+ uint32_t dis_opp_cas : 1; // Bits 0:0
+diff --git a/src/northbridge/intel/haswell/native_raminit/train_receive_enable.c b/src/northbridge/intel/haswell/native_raminit/train_receive_enable.c
+new file mode 100644
+index 0000000000..576c6bc21e
+--- /dev/null
++++ b/src/northbridge/intel/haswell/native_raminit/train_receive_enable.c
+@@ -0,0 +1,561 @@
++/* SPDX-License-Identifier: GPL-2.0-or-later */
++
++#include <console/console.h>
++#include <northbridge/intel/haswell/haswell.h>
++#include <types.h>
++
++#include "raminit_native.h"
++#include "ranges.h"
++
++#define RCVEN_PLOT RAM_DEBUG
++
++static enum raminit_status change_rcven_timing(struct sysinfo *ctrl, const uint8_t channel)
++{
++ int16_t max_rcven = -4096;
++ int16_t min_rcven = 4096;
++ int16_t max_rcven_rank[NUM_SLOTRANKS];
++ int16_t min_rcven_rank[NUM_SLOTRANKS];
++ for (uint8_t rank = 0; rank < NUM_SLOTRANKS; rank++) {
++ max_rcven_rank[rank] = max_rcven;
++ min_rcven_rank[rank] = min_rcven;
++ }
++ for (uint8_t rank = 0; rank < NUM_SLOTRANKS; rank++) {
++ if (!rank_in_ch(ctrl, rank, channel))
++ continue;
++
++ for (uint8_t byte = 0; byte < ctrl->lanes; byte++) {
++ int16_t new_rcven = ctrl->rcven[channel][rank][byte];
++ new_rcven -= ctrl->io_latency[channel][rank] * 64;
++ if (max_rcven_rank[rank] < new_rcven)
++ max_rcven_rank[rank] = new_rcven;
++
++ if (min_rcven_rank[rank] > new_rcven)
++ min_rcven_rank[rank] = new_rcven;
++ }
++ if (max_rcven < max_rcven_rank[rank])
++ max_rcven = max_rcven_rank[rank];
++
++ if (min_rcven > min_rcven_rank[rank])
++ min_rcven = min_rcven_rank[rank];
++ }
++
++ /*
++ * Determine how far we are from the ideal center point for RcvEn timing.
++ * (PiIdeal - AveRcvEn) / 64 is the ideal number of cycles we should have
++ * for IO latency. command training will reduce this by 64, so plan for
++ * that now in the ideal value. Round to closest integer.
++ */
++ const int16_t rre_pi_ideal = 256 + 64;
++ const int16_t pi_reserve = 64;
++ const int16_t rcven_center = (max_rcven + min_rcven) / 2;
++ const int8_t iolat_target = DIV_ROUND_CLOSEST(rre_pi_ideal - rcven_center, 64);
++
++ int8_t io_g_offset = 0;
++ int8_t io_lat[NUM_SLOTRANKS] = { 0 };
++ for (uint8_t rank = 0; rank < NUM_SLOTRANKS; rank++) {
++ if (!rank_in_ch(ctrl, rank, channel))
++ continue;
++
++ io_lat[rank] = iolat_target;
++
++ /* Check for RcvEn underflow/overflow */
++ const int16_t rcven_lower = 64 * io_lat[rank] + min_rcven_rank[rank];
++ if (rcven_lower < pi_reserve)
++ io_lat[rank] += DIV_ROUND_UP(pi_reserve - rcven_lower, 64);
++
++ const int16_t rcven_upper = 64 * io_lat[rank] + max_rcven_rank[rank];
++ if (rcven_upper > 511 - pi_reserve)
++ io_lat[rank] -= DIV_ROUND_UP(rcven_upper - (511 - pi_reserve), 64);
++
++ /* Check for IO latency over/underflow */
++ if (io_lat[rank] - io_g_offset > 14)
++ io_g_offset = io_lat[rank] - 14;
++
++ if (io_lat[rank] - io_g_offset < 1)
++ io_g_offset = io_lat[rank] - 1;
++
++ const int8_t cycle_offset = io_lat[rank] - ctrl->io_latency[channel][rank];
++ for (uint8_t byte = 0; byte < ctrl->lanes; byte++) {
++ ctrl->rcven[channel][rank][byte] += 64 * cycle_offset;
++ update_rxt(ctrl, channel, rank, byte, RXT_RESTORE, 0);
++ }
++ }
++
++ /* Calculate new IO comp latency */
++ union sc_io_latency_reg sc_io_lat = {
++ .raw = mchbar_read32(SC_IO_LATENCY_ch(channel)),
++ };
++
++ /* Check if we are underflowing or overflowing this field */
++ if (io_g_offset < 0 && sc_io_lat.rt_iocomp < -io_g_offset) {
++ printk(BIOS_ERR, "%s: IO COMP underflow\n", __func__);
++ printk(BIOS_ERR, "io_g_offset: %d\n", io_g_offset);
++ printk(BIOS_ERR, "rt_iocomp: %u\n", sc_io_lat.rt_iocomp);
++ return RAMINIT_STATUS_RCVEN_FAILURE;
++ }
++ if (io_g_offset > 0 && io_g_offset > 0x3f - sc_io_lat.rt_iocomp) {
++ printk(BIOS_ERR, "%s: IO COMP overflow\n", __func__);
++ printk(BIOS_ERR, "io_g_offset: %d\n", io_g_offset);
++ printk(BIOS_ERR, "rt_iocomp: %u\n", sc_io_lat.rt_iocomp);
++ return RAMINIT_STATUS_RCVEN_FAILURE;
++ }
++ sc_io_lat.rt_iocomp += io_g_offset;
++ ctrl->rt_io_comp[channel] = sc_io_lat.rt_iocomp;
++ for (uint8_t rank = 0; rank < NUM_SLOTRANKS; rank++) {
++ if (ctrl->rankmap[channel] & BIT(rank))
++ ctrl->io_latency[channel][rank] = io_lat[rank] - io_g_offset;
++
++ const uint8_t shift = rank * 4;
++ sc_io_lat.raw &= ~(0xf << shift);
++ sc_io_lat.raw |= ctrl->io_latency[channel][rank] << shift;
++ }
++ mchbar_write32(SC_IO_LATENCY_ch(channel), sc_io_lat.raw);
++ return RAMINIT_STATUS_SUCCESS;
++}
++
++#define RL_START (256 + 24)
++#define RL_STOP (384 + 24)
++#define RL_STEP 8
++
++#define RE_NUM_SAMPLES 6
++
++static enum raminit_status verify_high_region(const int32_t center, const int32_t lwidth)
++{
++ if (center > RL_STOP) {
++ /* Check if center of high was found where it should be */
++ printk(BIOS_ERR, "RcvEn: Center of high (%d) higher than expected\n", center);
++ return RAMINIT_STATUS_RCVEN_FAILURE;
++ }
++ if (lwidth <= 32) {
++ /* Check if width is large enough */
++ printk(BIOS_ERR, "RcvEn: Width of high region (%d) too small\n", lwidth);
++ return RAMINIT_STATUS_RCVEN_FAILURE;
++ }
++ if (lwidth >= 96) {
++ /* Since we're calibrating a phase, a too large region is a problem */
++ printk(BIOS_ERR, "RcvEn: Width of high region (%d) too large\n", lwidth);
++ return RAMINIT_STATUS_RCVEN_FAILURE;
++ }
++ return RAMINIT_STATUS_SUCCESS;
++}
++
++static void program_io_latency(struct sysinfo *ctrl, const uint8_t channel, const uint8_t rank)
++{
++ const uint8_t shift = rank * 4;
++ const uint8_t iolat = ctrl->io_latency[channel][rank];
++ mchbar_clrsetbits32(SC_IO_LATENCY_ch(channel), 0xf << shift, iolat << shift);
++}
++
++static void program_rl_delays(struct sysinfo *ctrl, const uint8_t rank, const uint16_t rl_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_rxt(ctrl, channel, rank, byte, RXT_RCVEN, rl_delay);
++ }
++}
++
++static bool sample_dqs(const uint8_t channel, const uint8_t byte)
++{
++ return (get_data_train_feedback(channel, byte) & 0x1ff) >= BIT(RE_NUM_SAMPLES - 1);
++}
++
++enum raminit_status train_receive_enable(struct sysinfo *ctrl)
++{
++ const struct reut_box reut_addr = {
++ .col = {
++ .start = 0,
++ .stop = 1023,
++ .inc_rate = 0,
++ .inc_val = 1,
++ },
++ };
++ const struct wdb_pat wdb_pattern = {
++ .start_ptr = 0,
++ .stop_ptr = 9,
++ .inc_rate = 32,
++ .dq_pattern = BASIC_VA,
++ };
++
++ const uint16_t bytemask = BIT(ctrl->lanes) - 1;
++ const uint8_t fine_step = 1;
++
++ const uint8_t rt_delta = is_hsw_ult() ? 4 : 2;
++ const uint8_t rt_io_comp = 21 + rt_delta;
++ const uint8_t rt_latency = 16 + rt_delta;
++ setup_io_test(
++ ctrl,
++ ctrl->chanmap,
++ PAT_RD,
++ 2,
++ RE_NUM_SAMPLES + 1,
++ &reut_addr,
++ 0,
++ &wdb_pattern,
++ 0,
++ 8);
++
++ for (uint8_t channel = 0; channel < NUM_CHANNELS; channel++) {
++ if (!does_ch_exist(ctrl, channel))
++ continue;
++
++ 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_rx_on = 1;
++ mchbar_write32(DQ_CONTROL_2(channel, byte), data_control_2.raw);
++ }
++ union ddr_data_control_0_reg data_control_0 = {
++ .raw = ctrl->dq_control_0[channel],
++ };
++ if (ctrl->lpddr) {
++ /**
++ * W/A for b4618574 - @todo: remove for HSW ULT C0
++ * Can't have force_odt_on together with leaker, disable LPDDR
++ * mode during this training step. lpddr_mode is restored
++ * at the end of this function from the host structure.
++ */
++ data_control_0.lpddr_mode = 0;
++ mchbar_write32(DDR_DATA_ch_CONTROL_0(channel), data_control_0.raw);
++ }
++ data_control_0.force_odt_on = 1;
++ data_control_0.rl_training_mode = 1;
++ mchbar_write32(DDR_DATA_ch_CONTROL_0(channel), data_control_0.raw);
++ mchbar_write32(SC_IO_LATENCY_ch(channel), (union sc_io_latency_reg) {
++ .rt_iocomp = rt_io_comp,
++ }.raw);
++ }
++ enum raminit_status status = RAMINIT_STATUS_SUCCESS;
++ for (uint8_t rank = 0; rank < NUM_SLOTRANKS; rank++) {
++ if (!does_rank_exist(ctrl, rank))
++ continue;
++
++ /*
++ * Set initial roundtrip latency values. Assume -4 QCLK for worst board
++ * layout. This is calculated as HW_ROUNDT_LAT_DEFAULT_VALUE plus:
++ *
++ * DDR3: Default + (2 * tAA) + 4 QCLK + PI_CLK + N-mode value * 2
++ * LPDDR3: Default + (2 * tAA) + 4 QCLK + PI_CLK + tDQSCK_max
++ *
++ * N-mode is 3 during training mode. Both channels use the same timings.
++ */
++ /** TODO: differs for LPDDR **/
++ const uint32_t tmp = MAX(ctrl->multiplier, 4) + 5 + 2 * ctrl->tAA;
++ const uint32_t initial_rt_latency = MIN(rt_latency + tmp, 0x3f);
++
++ uint8_t chanmask = 0;
++ for (uint8_t channel = 0; channel < NUM_CHANNELS; channel++) {
++ chanmask |= select_reut_ranks(ctrl, channel, BIT(rank));
++ if (!rank_in_ch(ctrl, rank, channel))
++ continue;
++
++ ctrl->io_latency[channel][rank] = 0;
++ mchbar_write8(SC_ROUNDT_LAT_ch(channel) + rank, initial_rt_latency);
++ ctrl->rt_latency[channel][rank] = initial_rt_latency;
++ }
++
++ printk(BIOS_DEBUG, "Rank %u\n", rank);
++ printk(BIOS_DEBUG, "Steps 1 and 2: Find middle of high region\n");
++ printk(RCVEN_PLOT, "Byte");
++ for (uint8_t channel = 0; channel < NUM_CHANNELS; channel++) {
++ if (!rank_in_ch(ctrl, rank, channel))
++ continue;
++
++ printk(RCVEN_PLOT, "\t");
++ for (uint8_t byte = 0; byte < ctrl->lanes; byte++)
++ printk(RCVEN_PLOT, "%u ", byte);
++ }
++ printk(RCVEN_PLOT, "\nRcvEn\n");
++ struct phase_train_data region_data[NUM_CHANNELS][NUM_LANES] = { 0 };
++ for (uint16_t rl_delay = RL_START; rl_delay < RL_STOP; rl_delay += RL_STEP) {
++ printk(RCVEN_PLOT, " % 3d", rl_delay);
++ program_rl_delays(ctrl, rank, rl_delay);
++ run_io_test(ctrl, chanmask, BASIC_VA, true);
++ for (uint8_t channel = 0; channel < NUM_CHANNELS; channel++) {
++ if (!rank_in_ch(ctrl, rank, channel))
++ continue;
++
++ printk(RCVEN_PLOT, "\t");
++ for (uint8_t byte = 0; byte < ctrl->lanes; byte++) {
++ const bool high = sample_dqs(channel, byte);
++ printk(RCVEN_PLOT, high ? ". " : "# ");
++ phase_record_pass(
++ &region_data[channel][byte],
++ high,
++ rl_delay,
++ RL_START,
++ RL_STEP);
++ }
++ }
++ printk(RCVEN_PLOT, "\n");
++ }
++ printk(RCVEN_PLOT, "\n");
++ printk(BIOS_DEBUG, "Update RcvEn timing to be in the center of high region\n");
++ for (uint8_t channel = 0; channel < NUM_CHANNELS; channel++) {
++ if (!rank_in_ch(ctrl, rank, channel))
++ continue;
++
++ printk(BIOS_DEBUG, "C%u.R%u: \tLeft\tRight\tWidth\tCenter\n",
++ channel, rank);
++ for (uint8_t byte = 0; byte < ctrl->lanes; byte++) {
++ struct phase_train_data *const curr_data =
++ &region_data[channel][byte];
++ phase_append_current_to_initial(curr_data, RL_START, RL_STEP);
++ const int32_t lwidth = range_width(curr_data->largest);
++ const int32_t center = range_center(curr_data->largest);
++ printk(BIOS_DEBUG, " B%u: \t%d\t%d\t%d\t%d\n",
++ byte,
++ curr_data->largest.start,
++ curr_data->largest.end,
++ lwidth,
++ center);
++
++ status = verify_high_region(center, lwidth);
++ if (status) {
++ printk(BIOS_ERR,
++ "RcvEn problems on channel %u, byte %u\n",
++ channel, byte);
++ goto clean_up;
++ }
++ ctrl->rcven[channel][rank][byte] = center;
++ update_rxt(ctrl, channel, rank, byte, RXT_RESTORE, 0);
++ }
++ printk(BIOS_DEBUG, "\n");
++ }
++
++ printk(BIOS_DEBUG, "Step 3: Quarter preamble - Walk backwards\n");
++ printk(RCVEN_PLOT, "Byte");
++ for (uint8_t channel = 0; channel < NUM_CHANNELS; channel++) {
++ if (!rank_in_ch(ctrl, rank, channel))
++ continue;
++
++ printk(RCVEN_PLOT, "\t");
++ for (uint8_t byte = 0; byte < ctrl->lanes; byte++)
++ printk(RCVEN_PLOT, "%u ", byte);
++ }
++ printk(RCVEN_PLOT, "\nIOLAT\n");
++ bool done = false;
++ while (!done) {
++ 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(RCVEN_PLOT, " %2u\t", ctrl->io_latency[channel][rank]);
++ uint16_t highs = 0;
++ for (uint8_t byte = 0; byte < ctrl->lanes; byte++) {
++ const bool high = sample_dqs(channel, byte);
++ printk(RCVEN_PLOT, high ? "H " : "L ");
++ if (high)
++ highs |= BIT(byte);
++ }
++ if (!highs)
++ continue;
++
++ done = false;
++
++ /* If all bytes sample high, adjust timing globally */
++ if (highs == bytemask && ctrl->io_latency[channel][rank] < 14) {
++ ctrl->io_latency[channel][rank] += 2;
++ ctrl->io_latency[channel][rank] %= 16;
++ program_io_latency(ctrl, channel, rank);
++ continue;
++ }
++
++ /* Otherwise, adjust individual bytes */
++ for (uint8_t byte = 0; byte < ctrl->lanes; byte++) {
++ if (!(highs & BIT(byte)))
++ continue;
++
++ if (ctrl->rcven[channel][rank][byte] < 128) {
++ printk(BIOS_ERR,
++ "RcvEn underflow: walking backwards\n");
++ printk(BIOS_ERR,
++ "For channel %u, rank %u, byte %u\n",
++ channel, rank, byte);
++ status = RAMINIT_STATUS_RCVEN_FAILURE;
++ goto clean_up;
++ }
++ ctrl->rcven[channel][rank][byte] -= 128;
++ update_rxt(ctrl, channel, rank, byte, RXT_RESTORE, 0);
++ }
++ }
++ printk(RCVEN_PLOT, "\n");
++ }
++ for (uint8_t channel = 0; channel < NUM_CHANNELS; channel++) {
++ if (!rank_in_ch(ctrl, rank, channel))
++ continue;
++
++ printk(BIOS_DEBUG, "\nC%u: Preamble\n", channel);
++ for (uint8_t byte = 0; byte < ctrl->lanes; byte++) {
++ printk(BIOS_DEBUG,
++ " B%u: %u\n", byte, ctrl->rcven[channel][rank][byte]);
++ }
++ }
++ printk(BIOS_DEBUG, "\n");
++
++ printk(BIOS_DEBUG, "Step 4: Add 1 qclk\n");
++ 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++) {
++ ctrl->rcven[channel][rank][byte] += 64;
++ update_rxt(ctrl, channel, rank, byte, RXT_RESTORE, 0);
++ }
++ }
++ printk(BIOS_DEBUG, "\n");
++
++ printk(BIOS_DEBUG, "Step 5: Walk forward to find rising edge\n");
++ printk(RCVEN_PLOT, "Byte");
++ for (uint8_t channel = 0; channel < NUM_CHANNELS; channel++) {
++ if (!rank_in_ch(ctrl, rank, channel))
++ continue;
++
++ printk(RCVEN_PLOT, "\t");
++ for (uint8_t byte = 0; byte < ctrl->lanes; byte++)
++ printk(RCVEN_PLOT, "%u ", byte);
++ }
++ printk(RCVEN_PLOT, "\n inc\n");
++ uint16_t ch_result[NUM_CHANNELS] = { 0 };
++ uint8_t inc_preamble[NUM_CHANNELS][NUM_LANES] = { 0 };
++ for (uint8_t inc = 0; inc < 64; inc += fine_step) {
++ printk(RCVEN_PLOT, " %2u\t", inc);
++ 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;
++
++ for (uint8_t byte = 0; byte < ctrl->lanes; byte++) {
++ if (ch_result[channel] & BIT(byte)) {
++ /* Skip bytes that are already done */
++ printk(RCVEN_PLOT, ". ");
++ continue;
++ }
++ const bool pass = sample_dqs(channel, byte);
++ printk(RCVEN_PLOT, pass ? ". " : "# ");
++ if (pass) {
++ ch_result[channel] |= BIT(byte);
++ continue;
++ }
++ ctrl->rcven[channel][rank][byte] += fine_step;
++ update_rxt(ctrl, channel, rank, byte, RXT_RESTORE, 0);
++ inc_preamble[channel][byte] = inc;
++ }
++ printk(RCVEN_PLOT, "\t");
++ if (ch_result[channel] != bytemask)
++ done = false;
++ }
++ printk(RCVEN_PLOT, "\n");
++ if (done)
++ break;
++ }
++ printk(BIOS_DEBUG, "\n");
++ if (!done) {
++ printk(BIOS_ERR, "Error: Preamble edge not found for all bytes\n");
++ printk(BIOS_ERR, "The final RcvEn results are as follows:\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: preamble\n",
++ channel, rank);
++ for (uint8_t byte = 0; byte < ctrl->lanes; byte++) {
++ printk(BIOS_ERR, " Byte %u: %u%s\n", byte,
++ ctrl->rcven[channel][rank][byte],
++ (ch_result[channel] ^ bytemask) & BIT(byte)
++ ? ""
++ : " *** Check this byte! ***");
++ }
++ }
++ status = RAMINIT_STATUS_RCVEN_FAILURE;
++ goto clean_up;
++ }
++
++ printk(BIOS_DEBUG, "Step 6: center on preamble and clean up rank\n");
++ for (uint8_t channel = 0; channel < NUM_CHANNELS; channel++) {
++ if (!rank_in_ch(ctrl, rank, channel))
++ continue;
++
++ printk(BIOS_DEBUG, "C%u: Preamble increment\n", channel);
++ for (uint8_t byte = 0; byte < ctrl->lanes; byte++) {
++ /*
++ * For Traditional, pull in RcvEn by 64. For ULT, take the DQS
++ * drift into account to the specified guardband: tDQSCK_DRIFT.
++ */
++ ctrl->rcven[channel][rank][byte] -= tDQSCK_DRIFT;
++ update_rxt(ctrl, channel, rank, byte, RXT_RESTORE, 0);
++ printk(BIOS_DEBUG, " B%u: %u %u\n", byte,
++ ctrl->rcven[channel][rank][byte],
++ inc_preamble[channel][byte]);
++ }
++ printk(BIOS_DEBUG, "\n");
++ }
++ printk(BIOS_DEBUG, "\n");
++ }
++
++clean_up:
++ for (uint8_t channel = 0; channel < NUM_CHANNELS; channel++) {
++ if (!does_ch_exist(ctrl, channel))
++ continue;
++
++ if (ctrl->lpddr) {
++ /**
++ * W/A for b4618574 - @todo: remove for HSW ULT C0
++ * Can't have force_odt_on together with leaker, disable LPDDR mode for
++ * this training step. This write will disable force_odt_on while still
++ * keeping LPDDR mode disabled. Second write will restore LPDDR mode.
++ */
++ union ddr_data_control_0_reg data_control_0 = {
++ .raw = ctrl->dq_control_0[channel],
++ };
++ data_control_0.lpddr_mode = 0;
++ mchbar_write32(DDR_DATA_ch_CONTROL_0(channel), data_control_0.raw);
++ }
++ 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]);
++ }
++ }
++ io_reset();
++ if (status)
++ return status;
++
++ printk(BIOS_DEBUG, "Step 7: Sync IO latency across all ranks\n");
++ for (uint8_t channel = 0; channel < NUM_CHANNELS; channel++) {
++ if (!does_ch_exist(ctrl, channel))
++ continue;
++
++ status = change_rcven_timing(ctrl, channel);
++ if (status)
++ return status;
++ }
++ printk(BIOS_DEBUG, "\nFinal Receive Enable and IO latency settings:\n");
++ for (uint8_t channel = 0; channel < NUM_CHANNELS; channel++) {
++ if (!does_ch_exist(ctrl, channel))
++ continue;
++
++ for (uint8_t rank = 0; rank < NUM_SLOTRANKS; rank++) {
++ if (!rank_in_ch(ctrl, rank, channel))
++ continue;
++
++ const union sc_io_latency_reg sc_io_latency = {
++ .raw = mchbar_read32(SC_IO_LATENCY_ch(channel)),
++ };
++ printk(BIOS_DEBUG, " C%u.R%u: IOLAT = %u rt_iocomp = %u\n", channel,
++ rank, ctrl->io_latency[channel][rank], sc_io_latency.rt_iocomp);
++ for (uint8_t byte = 0; byte < ctrl->lanes; byte++) {
++ printk(BIOS_DEBUG, " B%u: %u\n", byte,
++ ctrl->rcven[channel][rank][byte]);
++ }
++ printk(BIOS_DEBUG, "\n");
++ }
++ }
++ return status;
++}
+diff --git a/src/northbridge/intel/haswell/registers/mchbar.h b/src/northbridge/intel/haswell/registers/mchbar.h
+index a81559bb1e..9172d4f2b0 100644
+--- a/src/northbridge/intel/haswell/registers/mchbar.h
++++ b/src/northbridge/intel/haswell/registers/mchbar.h
+@@ -18,6 +18,8 @@
+ #define RX_TRAIN_ch_r_b(ch, rank, byte) _DDRIO_C_R_B(0x0000, ch, rank, byte)
+ #define TX_TRAIN_ch_r_b(ch, rank, byte) _DDRIO_C_R_B(0x0020, ch, rank, byte)
+
++#define DDR_DATA_TRAIN_FEEDBACK(ch, byte) _DDRIO_C_R_B(0x0054, ch, 0, byte)
++
+ #define DQ_CONTROL_2(ch, byte) _DDRIO_C_R_B(0x0064, ch, 0, byte)
+ #define DQ_CONTROL_0(ch, byte) _DDRIO_C_R_B(0x0074, ch, 0, byte)
+
+@@ -100,6 +102,7 @@
+ #define COMMAND_RATE_LIMIT_ch(ch) _MCMAIN_C(0x4010, ch)
+ #define TC_BANK_RANK_D_ch(ch) _MCMAIN_C(0x4014, ch)
+ #define SC_ROUNDT_LAT_ch(ch) _MCMAIN_C(0x4024, ch)
++#define SC_IO_LATENCY_ch(ch) _MCMAIN_C(0x4028, ch)
+
+ #define REUT_ch_PAT_WDB_CL_MUX_CFG(ch) _MCMAIN_C(0x4040, ch)
+
+--
+2.39.2
+
generated by cgit v1.2.1 (git 2.18.0) at 2024-11-09 18:47:40 +0000