2 * Copyright Altera Corporation (C) 2012-2015
4 * SPDX-License-Identifier: BSD-3-Clause
9 #include <asm/arch/sdram.h>
10 #include "sequencer.h"
11 #include "sequencer_auto.h"
12 #include "sequencer_auto_ac_init.h"
13 #include "sequencer_auto_inst_init.h"
14 #include "sequencer_defines.h"
16 static struct socfpga_sdr_rw_load_manager *sdr_rw_load_mgr_regs =
17 (struct socfpga_sdr_rw_load_manager *)(SDR_PHYGRP_RWMGRGRP_ADDRESS | 0x800);
19 static struct socfpga_sdr_rw_load_jump_manager *sdr_rw_load_jump_mgr_regs =
20 (struct socfpga_sdr_rw_load_jump_manager *)(SDR_PHYGRP_RWMGRGRP_ADDRESS | 0xC00);
22 static struct socfpga_sdr_reg_file *sdr_reg_file =
23 (struct socfpga_sdr_reg_file *)SDR_PHYGRP_REGFILEGRP_ADDRESS;
25 static struct socfpga_sdr_scc_mgr *sdr_scc_mgr =
26 (struct socfpga_sdr_scc_mgr *)(SDR_PHYGRP_SCCGRP_ADDRESS | 0xe00);
28 static struct socfpga_phy_mgr_cmd *phy_mgr_cmd =
29 (struct socfpga_phy_mgr_cmd *)SDR_PHYGRP_PHYMGRGRP_ADDRESS;
31 static struct socfpga_phy_mgr_cfg *phy_mgr_cfg =
32 (struct socfpga_phy_mgr_cfg *)(SDR_PHYGRP_PHYMGRGRP_ADDRESS | 0x40);
34 static struct socfpga_data_mgr *data_mgr =
35 (struct socfpga_data_mgr *)SDR_PHYGRP_DATAMGRGRP_ADDRESS;
37 static struct socfpga_sdr_ctrl *sdr_ctrl =
38 (struct socfpga_sdr_ctrl *)SDR_CTRLGRP_ADDRESS;
43 * In order to reduce ROM size, most of the selectable calibration steps are
44 * decided at compile time based on the user's calibration mode selection,
45 * as captured by the STATIC_CALIB_STEPS selection below.
47 * However, to support simulation-time selection of fast simulation mode, where
48 * we skip everything except the bare minimum, we need a few of the steps to
49 * be dynamic. In those cases, we either use the DYNAMIC_CALIB_STEPS for the
50 * check, which is based on the rtl-supplied value, or we dynamically compute
51 * the value to use based on the dynamically-chosen calibration mode
55 #define STATIC_IN_RTL_SIM 0
56 #define STATIC_SKIP_DELAY_LOOPS 0
58 #define STATIC_CALIB_STEPS (STATIC_IN_RTL_SIM | CALIB_SKIP_FULL_TEST | \
59 STATIC_SKIP_DELAY_LOOPS)
61 /* calibration steps requested by the rtl */
62 uint16_t dyn_calib_steps;
65 * To make CALIB_SKIP_DELAY_LOOPS a dynamic conditional option
66 * instead of static, we use boolean logic to select between
67 * non-skip and skip values
69 * The mask is set to include all bits when not-skipping, but is
73 uint16_t skip_delay_mask; /* mask off bits when skipping/not-skipping */
75 #define SKIP_DELAY_LOOP_VALUE_OR_ZERO(non_skip_value) \
76 ((non_skip_value) & skip_delay_mask)
79 struct param_type *param;
80 uint32_t curr_shadow_reg;
82 static uint32_t rw_mgr_mem_calibrate_write_test(uint32_t rank_bgn,
83 uint32_t write_group, uint32_t use_dm,
84 uint32_t all_correct, uint32_t *bit_chk, uint32_t all_ranks);
86 static void set_failing_group_stage(uint32_t group, uint32_t stage,
90 * Only set the global stage if there was not been any other
93 if (gbl->error_stage == CAL_STAGE_NIL) {
94 gbl->error_substage = substage;
95 gbl->error_stage = stage;
96 gbl->error_group = group;
100 static void reg_file_set_group(u16 set_group)
102 clrsetbits_le32(&sdr_reg_file->cur_stage, 0xffff0000, set_group << 16);
105 static void reg_file_set_stage(u8 set_stage)
107 clrsetbits_le32(&sdr_reg_file->cur_stage, 0xffff, set_stage & 0xff);
110 static void reg_file_set_sub_stage(u8 set_sub_stage)
112 set_sub_stage &= 0xff;
113 clrsetbits_le32(&sdr_reg_file->cur_stage, 0xff00, set_sub_stage << 8);
116 static void initialize(void)
118 debug("%s:%d\n", __func__, __LINE__);
119 /* USER calibration has control over path to memory */
121 * In Hard PHY this is a 2-bit control:
125 writel(0x3, &phy_mgr_cfg->mux_sel);
127 /* USER memory clock is not stable we begin initialization */
128 writel(0, &phy_mgr_cfg->reset_mem_stbl);
130 /* USER calibration status all set to zero */
131 writel(0, &phy_mgr_cfg->cal_status);
133 writel(0, &phy_mgr_cfg->cal_debug_info);
135 if ((dyn_calib_steps & CALIB_SKIP_ALL) != CALIB_SKIP_ALL) {
136 param->read_correct_mask_vg = ((uint32_t)1 <<
137 (RW_MGR_MEM_DQ_PER_READ_DQS /
138 RW_MGR_MEM_VIRTUAL_GROUPS_PER_READ_DQS)) - 1;
139 param->write_correct_mask_vg = ((uint32_t)1 <<
140 (RW_MGR_MEM_DQ_PER_READ_DQS /
141 RW_MGR_MEM_VIRTUAL_GROUPS_PER_READ_DQS)) - 1;
142 param->read_correct_mask = ((uint32_t)1 <<
143 RW_MGR_MEM_DQ_PER_READ_DQS) - 1;
144 param->write_correct_mask = ((uint32_t)1 <<
145 RW_MGR_MEM_DQ_PER_WRITE_DQS) - 1;
146 param->dm_correct_mask = ((uint32_t)1 <<
147 (RW_MGR_MEM_DATA_WIDTH / RW_MGR_MEM_DATA_MASK_WIDTH))
152 static void set_rank_and_odt_mask(uint32_t rank, uint32_t odt_mode)
154 uint32_t odt_mask_0 = 0;
155 uint32_t odt_mask_1 = 0;
156 uint32_t cs_and_odt_mask;
158 if (odt_mode == RW_MGR_ODT_MODE_READ_WRITE) {
159 if (RW_MGR_MEM_NUMBER_OF_RANKS == 1) {
167 } else if (RW_MGR_MEM_NUMBER_OF_RANKS == 2) {
169 if (RW_MGR_MEM_NUMBER_OF_CS_PER_DIMM == 1) {
170 /* - Dual-Slot , Single-Rank
171 * (1 chip-select per DIMM)
173 * - RDIMM, 4 total CS (2 CS per DIMM)
175 * Since MEM_NUMBER_OF_RANKS is 2 they are
177 * with 2 CS each (special for RDIMM)
178 * Read: Turn on ODT on the opposite rank
179 * Write: Turn on ODT on all ranks
181 odt_mask_0 = 0x3 & ~(1 << rank);
185 * USER - Single-Slot , Dual-rank DIMMs
186 * (2 chip-selects per DIMM)
187 * USER Read: Turn on ODT off on all ranks
188 * USER Write: Turn on ODT on active rank
191 odt_mask_1 = 0x3 & (1 << rank);
196 * ----------+-----------------------+
199 * Read From +-----------------------+
200 * Rank | 3 | 2 | 1 | 0 |
201 * ----------+-----+-----+-----+-----+
202 * 0 | 0 | 1 | 0 | 0 |
203 * 1 | 1 | 0 | 0 | 0 |
204 * 2 | 0 | 0 | 0 | 1 |
205 * 3 | 0 | 0 | 1 | 0 |
206 * ----------+-----+-----+-----+-----+
209 * ----------+-----------------------+
212 * Write To +-----------------------+
213 * Rank | 3 | 2 | 1 | 0 |
214 * ----------+-----+-----+-----+-----+
215 * 0 | 0 | 1 | 0 | 1 |
216 * 1 | 1 | 0 | 1 | 0 |
217 * 2 | 0 | 1 | 0 | 1 |
218 * 3 | 1 | 0 | 1 | 0 |
219 * ----------+-----+-----+-----+-----+
246 (0xFF & ~(1 << rank)) |
247 ((0xFF & odt_mask_0) << 8) |
248 ((0xFF & odt_mask_1) << 16);
249 writel(cs_and_odt_mask, SDR_PHYGRP_RWMGRGRP_ADDRESS |
250 RW_MGR_SET_CS_AND_ODT_MASK_OFFSET);
254 * scc_mgr_set() - Set SCC Manager register
255 * @off: Base offset in SCC Manager space
256 * @grp: Read/Write group
257 * @val: Value to be set
259 * This function sets the SCC Manager (Scan Chain Control Manager) register.
261 static void scc_mgr_set(u32 off, u32 grp, u32 val)
263 writel(val, SDR_PHYGRP_SCCGRP_ADDRESS | off | (grp << 2));
267 * scc_mgr_initialize() - Initialize SCC Manager registers
269 * Initialize SCC Manager registers.
271 static void scc_mgr_initialize(void)
274 * Clear register file for HPS. 16 (2^4) is the size of the
275 * full register file in the scc mgr:
276 * RFILE_DEPTH = 1 + log2(MEM_DQ_PER_DQS + 1 + MEM_DM_PER_DQS +
277 * MEM_IF_READ_DQS_WIDTH - 1);
281 for (i = 0; i < 16; i++) {
282 debug_cond(DLEVEL == 1, "%s:%d: Clearing SCC RFILE index %u\n",
283 __func__, __LINE__, i);
284 scc_mgr_set(SCC_MGR_HHP_RFILE_OFFSET, 0, i);
288 static void scc_mgr_set_dqdqs_output_phase(uint32_t write_group, uint32_t phase)
290 scc_mgr_set(SCC_MGR_DQDQS_OUT_PHASE_OFFSET, write_group, phase);
293 static void scc_mgr_set_dqs_bus_in_delay(uint32_t read_group, uint32_t delay)
295 scc_mgr_set(SCC_MGR_DQS_IN_DELAY_OFFSET, read_group, delay);
298 static void scc_mgr_set_dqs_en_phase(uint32_t read_group, uint32_t phase)
300 scc_mgr_set(SCC_MGR_DQS_EN_PHASE_OFFSET, read_group, phase);
303 static void scc_mgr_set_dqs_en_delay(uint32_t read_group, uint32_t delay)
305 scc_mgr_set(SCC_MGR_DQS_EN_DELAY_OFFSET, read_group, delay);
308 static void scc_mgr_set_dqs_io_in_delay(uint32_t write_group, uint32_t delay)
310 scc_mgr_set(SCC_MGR_IO_IN_DELAY_OFFSET, RW_MGR_MEM_DQ_PER_WRITE_DQS,
314 static void scc_mgr_set_dq_in_delay(uint32_t dq_in_group, uint32_t delay)
316 scc_mgr_set(SCC_MGR_IO_IN_DELAY_OFFSET, dq_in_group, delay);
319 static void scc_mgr_set_dq_out1_delay(uint32_t dq_in_group, uint32_t delay)
321 scc_mgr_set(SCC_MGR_IO_OUT1_DELAY_OFFSET, dq_in_group, delay);
324 static void scc_mgr_set_dqs_out1_delay(uint32_t write_group,
327 scc_mgr_set(SCC_MGR_IO_OUT1_DELAY_OFFSET, RW_MGR_MEM_DQ_PER_WRITE_DQS,
331 static void scc_mgr_set_dm_out1_delay(uint32_t dm, uint32_t delay)
333 scc_mgr_set(SCC_MGR_IO_OUT1_DELAY_OFFSET,
334 RW_MGR_MEM_DQ_PER_WRITE_DQS + 1 + dm,
338 /* load up dqs config settings */
339 static void scc_mgr_load_dqs(uint32_t dqs)
341 writel(dqs, &sdr_scc_mgr->dqs_ena);
344 /* load up dqs io config settings */
345 static void scc_mgr_load_dqs_io(void)
347 writel(0, &sdr_scc_mgr->dqs_io_ena);
350 /* load up dq config settings */
351 static void scc_mgr_load_dq(uint32_t dq_in_group)
353 writel(dq_in_group, &sdr_scc_mgr->dq_ena);
356 /* load up dm config settings */
357 static void scc_mgr_load_dm(uint32_t dm)
359 writel(dm, &sdr_scc_mgr->dm_ena);
363 * scc_mgr_set_all_ranks() - Set SCC Manager register for all ranks
364 * @off: Base offset in SCC Manager space
365 * @grp: Read/Write group
366 * @val: Value to be set
367 * @update: If non-zero, trigger SCC Manager update for all ranks
369 * This function sets the SCC Manager (Scan Chain Control Manager) register
370 * and optionally triggers the SCC update for all ranks.
372 static void scc_mgr_set_all_ranks(const u32 off, const u32 grp, const u32 val,
377 for (r = 0; r < RW_MGR_MEM_NUMBER_OF_RANKS;
378 r += NUM_RANKS_PER_SHADOW_REG) {
379 scc_mgr_set(off, grp, val);
381 if (update || (r == 0)) {
382 writel(grp, &sdr_scc_mgr->dqs_ena);
383 writel(0, &sdr_scc_mgr->update);
388 static void scc_mgr_set_dqs_en_phase_all_ranks(u32 read_group, u32 phase)
391 * USER although the h/w doesn't support different phases per
392 * shadow register, for simplicity our scc manager modeling
393 * keeps different phase settings per shadow reg, and it's
394 * important for us to keep them in sync to match h/w.
395 * for efficiency, the scan chain update should occur only
398 scc_mgr_set_all_ranks(SCC_MGR_DQS_EN_PHASE_OFFSET,
399 read_group, phase, 0);
402 static void scc_mgr_set_dqdqs_output_phase_all_ranks(uint32_t write_group,
406 * USER although the h/w doesn't support different phases per
407 * shadow register, for simplicity our scc manager modeling
408 * keeps different phase settings per shadow reg, and it's
409 * important for us to keep them in sync to match h/w.
410 * for efficiency, the scan chain update should occur only
413 scc_mgr_set_all_ranks(SCC_MGR_DQDQS_OUT_PHASE_OFFSET,
414 write_group, phase, 0);
417 static void scc_mgr_set_dqs_en_delay_all_ranks(uint32_t read_group,
421 * In shadow register mode, the T11 settings are stored in
422 * registers in the core, which are updated by the DQS_ENA
423 * signals. Not issuing the SCC_MGR_UPD command allows us to
424 * save lots of rank switching overhead, by calling
425 * select_shadow_regs_for_update with update_scan_chains
428 scc_mgr_set_all_ranks(SCC_MGR_DQS_EN_DELAY_OFFSET,
429 read_group, delay, 1);
430 writel(0, &sdr_scc_mgr->update);
434 * scc_mgr_set_oct_out1_delay() - Set OCT output delay
435 * @write_group: Write group
436 * @delay: Delay value
438 * This function sets the OCT output delay in SCC manager.
440 static void scc_mgr_set_oct_out1_delay(const u32 write_group, const u32 delay)
442 const int ratio = RW_MGR_MEM_IF_READ_DQS_WIDTH /
443 RW_MGR_MEM_IF_WRITE_DQS_WIDTH;
444 const int base = write_group * ratio;
447 * Load the setting in the SCC manager
448 * Although OCT affects only write data, the OCT delay is controlled
449 * by the DQS logic block which is instantiated once per read group.
450 * For protocols where a write group consists of multiple read groups,
451 * the setting must be set multiple times.
453 for (i = 0; i < ratio; i++)
454 scc_mgr_set(SCC_MGR_OCT_OUT1_DELAY_OFFSET, base + i, delay);
457 static void scc_mgr_set_hhp_extras(void)
460 * Load the fixed setting in the SCC manager
461 * bits: 0:0 = 1'b1 - dqs bypass
462 * bits: 1:1 = 1'b1 - dq bypass
463 * bits: 4:2 = 3'b001 - rfifo_mode
464 * bits: 6:5 = 2'b01 - rfifo clock_select
465 * bits: 7:7 = 1'b0 - separate gating from ungating setting
466 * bits: 8:8 = 1'b0 - separate OE from Output delay setting
468 uint32_t value = (0<<8) | (0<<7) | (1<<5) | (1<<2) | (1<<1) | (1<<0);
469 uint32_t addr = SDR_PHYGRP_SCCGRP_ADDRESS | SCC_MGR_HHP_GLOBALS_OFFSET;
471 writel(value, addr + SCC_MGR_HHP_EXTRAS_OFFSET);
475 * USER Zero all DQS config
476 * TODO: maybe rename to scc_mgr_zero_dqs_config (or something)
478 static void scc_mgr_zero_all(void)
483 * USER Zero all DQS config settings, across all groups and all
486 for (r = 0; r < RW_MGR_MEM_NUMBER_OF_RANKS; r +=
487 NUM_RANKS_PER_SHADOW_REG) {
488 for (i = 0; i < RW_MGR_MEM_IF_READ_DQS_WIDTH; i++) {
490 * The phases actually don't exist on a per-rank basis,
491 * but there's no harm updating them several times, so
492 * let's keep the code simple.
494 scc_mgr_set_dqs_bus_in_delay(i, IO_DQS_IN_RESERVE);
495 scc_mgr_set_dqs_en_phase(i, 0);
496 scc_mgr_set_dqs_en_delay(i, 0);
499 for (i = 0; i < RW_MGR_MEM_IF_WRITE_DQS_WIDTH; i++) {
500 scc_mgr_set_dqdqs_output_phase(i, 0);
501 /* av/cv don't have out2 */
502 scc_mgr_set_oct_out1_delay(i, IO_DQS_OUT_RESERVE);
506 /* multicast to all DQS group enables */
507 writel(0xff, &sdr_scc_mgr->dqs_ena);
508 writel(0, &sdr_scc_mgr->update);
512 * scc_set_bypass_mode() - Set bypass mode and trigger SCC update
513 * @write_group: Write group
515 * Set bypass mode and trigger SCC update.
517 static void scc_set_bypass_mode(const u32 write_group)
519 /* Only needed once to set all groups, pins, DQ, DQS, DM. */
520 if (write_group == 0) {
521 debug_cond(DLEVEL == 1, "%s:%d Setting HHP Extras\n", __func__,
523 scc_mgr_set_hhp_extras();
524 debug_cond(DLEVEL == 1, "%s:%d Done Setting HHP Extras\n",
528 /* Multicast to all DQ enables. */
529 writel(0xff, &sdr_scc_mgr->dq_ena);
530 writel(0xff, &sdr_scc_mgr->dm_ena);
532 /* Update current DQS IO enable. */
533 writel(0, &sdr_scc_mgr->dqs_io_ena);
535 /* Update the DQS logic. */
536 writel(write_group, &sdr_scc_mgr->dqs_ena);
539 writel(0, &sdr_scc_mgr->update);
543 * scc_mgr_load_dqs_for_write_group() - Load DQS settings for Write Group
544 * @write_group: Write group
546 * Load DQS settings for Write Group, do not trigger SCC update.
548 static void scc_mgr_load_dqs_for_write_group(const u32 write_group)
550 const int ratio = RW_MGR_MEM_IF_READ_DQS_WIDTH /
551 RW_MGR_MEM_IF_WRITE_DQS_WIDTH;
552 const int base = write_group * ratio;
555 * Load the setting in the SCC manager
556 * Although OCT affects only write data, the OCT delay is controlled
557 * by the DQS logic block which is instantiated once per read group.
558 * For protocols where a write group consists of multiple read groups,
559 * the setting must be set multiple times.
561 for (i = 0; i < ratio; i++)
562 writel(base + i, &sdr_scc_mgr->dqs_ena);
565 static void scc_mgr_zero_group(uint32_t write_group, uint32_t test_begin,
570 for (r = 0; r < RW_MGR_MEM_NUMBER_OF_RANKS; r +=
571 NUM_RANKS_PER_SHADOW_REG) {
572 /* Zero all DQ config settings */
573 for (i = 0; i < RW_MGR_MEM_DQ_PER_WRITE_DQS; i++) {
574 scc_mgr_set_dq_out1_delay(i, 0);
576 scc_mgr_set_dq_in_delay(i, 0);
579 /* multicast to all DQ enables */
580 writel(0xff, &sdr_scc_mgr->dq_ena);
582 /* Zero all DM config settings */
583 for (i = 0; i < RW_MGR_NUM_DM_PER_WRITE_GROUP; i++) {
584 scc_mgr_set_dm_out1_delay(i, 0);
587 /* multicast to all DM enables */
588 writel(0xff, &sdr_scc_mgr->dm_ena);
590 /* zero all DQS io settings */
592 scc_mgr_set_dqs_io_in_delay(write_group, 0);
593 /* av/cv don't have out2 */
594 scc_mgr_set_dqs_out1_delay(write_group, IO_DQS_OUT_RESERVE);
595 scc_mgr_set_oct_out1_delay(write_group, IO_DQS_OUT_RESERVE);
596 scc_mgr_load_dqs_for_write_group(write_group);
598 /* multicast to all DQS IO enables (only 1) */
599 writel(0, &sdr_scc_mgr->dqs_io_ena);
601 /* hit update to zero everything */
602 writel(0, &sdr_scc_mgr->update);
607 * apply and load a particular input delay for the DQ pins in a group
608 * group_bgn is the index of the first dq pin (in the write group)
610 static void scc_mgr_apply_group_dq_in_delay(uint32_t write_group,
611 uint32_t group_bgn, uint32_t delay)
615 for (i = 0, p = group_bgn; i < RW_MGR_MEM_DQ_PER_READ_DQS; i++, p++) {
616 scc_mgr_set_dq_in_delay(p, delay);
621 /* apply and load a particular output delay for the DQ pins in a group */
622 static void scc_mgr_apply_group_dq_out1_delay(uint32_t write_group,
628 for (i = 0, p = group_bgn; i < RW_MGR_MEM_DQ_PER_WRITE_DQS; i++, p++) {
629 scc_mgr_set_dq_out1_delay(i, delay1);
634 /* apply and load a particular output delay for the DM pins in a group */
635 static void scc_mgr_apply_group_dm_out1_delay(uint32_t write_group,
640 for (i = 0; i < RW_MGR_NUM_DM_PER_WRITE_GROUP; i++) {
641 scc_mgr_set_dm_out1_delay(i, delay1);
647 /* apply and load delay on both DQS and OCT out1 */
648 static void scc_mgr_apply_group_dqs_io_and_oct_out1(uint32_t write_group,
651 scc_mgr_set_dqs_out1_delay(write_group, delay);
652 scc_mgr_load_dqs_io();
654 scc_mgr_set_oct_out1_delay(write_group, delay);
655 scc_mgr_load_dqs_for_write_group(write_group);
658 /* apply a delay to the entire output side: DQ, DM, DQS, OCT */
659 static void scc_mgr_apply_group_all_out_delay_add(uint32_t write_group,
663 uint32_t i, p, new_delay;
666 for (i = 0, p = group_bgn; i < RW_MGR_MEM_DQ_PER_WRITE_DQS; i++, p++) {
667 new_delay = READ_SCC_DQ_OUT2_DELAY;
670 if (new_delay > IO_IO_OUT2_DELAY_MAX) {
671 debug_cond(DLEVEL == 1, "%s:%d (%u, %u, %u) DQ[%u,%u]:\
672 %u > %lu => %lu", __func__, __LINE__,
673 write_group, group_bgn, delay, i, p, new_delay,
674 (long unsigned int)IO_IO_OUT2_DELAY_MAX,
675 (long unsigned int)IO_IO_OUT2_DELAY_MAX);
676 new_delay = IO_IO_OUT2_DELAY_MAX;
683 for (i = 0; i < RW_MGR_NUM_DM_PER_WRITE_GROUP; i++) {
684 new_delay = READ_SCC_DM_IO_OUT2_DELAY;
687 if (new_delay > IO_IO_OUT2_DELAY_MAX) {
688 debug_cond(DLEVEL == 1, "%s:%d (%u, %u, %u) DM[%u]:\
689 %u > %lu => %lu\n", __func__, __LINE__,
690 write_group, group_bgn, delay, i, new_delay,
691 (long unsigned int)IO_IO_OUT2_DELAY_MAX,
692 (long unsigned int)IO_IO_OUT2_DELAY_MAX);
693 new_delay = IO_IO_OUT2_DELAY_MAX;
700 new_delay = READ_SCC_DQS_IO_OUT2_DELAY;
703 if (new_delay > IO_IO_OUT2_DELAY_MAX) {
704 debug_cond(DLEVEL == 1, "%s:%d (%u, %u, %u) DQS: %u > %d => %d;"
705 " adding %u to OUT1\n", __func__, __LINE__,
706 write_group, group_bgn, delay, new_delay,
707 IO_IO_OUT2_DELAY_MAX, IO_IO_OUT2_DELAY_MAX,
708 new_delay - IO_IO_OUT2_DELAY_MAX);
709 scc_mgr_set_dqs_out1_delay(write_group, new_delay -
710 IO_IO_OUT2_DELAY_MAX);
711 new_delay = IO_IO_OUT2_DELAY_MAX;
714 scc_mgr_load_dqs_io();
717 new_delay = READ_SCC_OCT_OUT2_DELAY;
720 if (new_delay > IO_IO_OUT2_DELAY_MAX) {
721 debug_cond(DLEVEL == 1, "%s:%d (%u, %u, %u) DQS: %u > %d => %d;"
722 " adding %u to OUT1\n", __func__, __LINE__,
723 write_group, group_bgn, delay, new_delay,
724 IO_IO_OUT2_DELAY_MAX, IO_IO_OUT2_DELAY_MAX,
725 new_delay - IO_IO_OUT2_DELAY_MAX);
726 scc_mgr_set_oct_out1_delay(write_group, new_delay -
727 IO_IO_OUT2_DELAY_MAX);
728 new_delay = IO_IO_OUT2_DELAY_MAX;
731 scc_mgr_load_dqs_for_write_group(write_group);
735 * USER apply a delay to the entire output side (DQ, DM, DQS, OCT)
738 static void scc_mgr_apply_group_all_out_delay_add_all_ranks(
739 uint32_t write_group, uint32_t group_bgn, uint32_t delay)
743 for (r = 0; r < RW_MGR_MEM_NUMBER_OF_RANKS;
744 r += NUM_RANKS_PER_SHADOW_REG) {
745 scc_mgr_apply_group_all_out_delay_add(write_group,
747 writel(0, &sdr_scc_mgr->update);
751 /* optimization used to recover some slots in ddr3 inst_rom */
752 /* could be applied to other protocols if we wanted to */
753 static void set_jump_as_return(void)
756 * to save space, we replace return with jump to special shared
757 * RETURN instruction so we set the counter to large value so that
760 writel(0xff, &sdr_rw_load_mgr_regs->load_cntr0);
761 writel(RW_MGR_RETURN, &sdr_rw_load_jump_mgr_regs->load_jump_add0);
765 * should always use constants as argument to ensure all computations are
766 * performed at compile time
768 static void delay_for_n_mem_clocks(const uint32_t clocks)
775 debug("%s:%d: clocks=%u ... start\n", __func__, __LINE__, clocks);
778 afi_clocks = (clocks + AFI_RATE_RATIO-1) / AFI_RATE_RATIO;
779 /* scale (rounding up) to get afi clocks */
782 * Note, we don't bother accounting for being off a little bit
783 * because of a few extra instructions in outer loops
784 * Note, the loops have a test at the end, and do the test before
785 * the decrement, and so always perform the loop
786 * 1 time more than the counter value
788 if (afi_clocks == 0) {
790 } else if (afi_clocks <= 0x100) {
791 inner = afi_clocks-1;
794 } else if (afi_clocks <= 0x10000) {
796 outer = (afi_clocks-1) >> 8;
801 c_loop = (afi_clocks-1) >> 16;
805 * rom instructions are structured as follows:
807 * IDLE_LOOP2: jnz cntr0, TARGET_A
808 * IDLE_LOOP1: jnz cntr1, TARGET_B
811 * so, when doing nested loops, TARGET_A is set to IDLE_LOOP2, and
812 * TARGET_B is set to IDLE_LOOP2 as well
814 * if we have no outer loop, though, then we can use IDLE_LOOP1 only,
815 * and set TARGET_B to IDLE_LOOP1 and we skip IDLE_LOOP2 entirely
817 * a little confusing, but it helps save precious space in the inst_rom
818 * and sequencer rom and keeps the delays more accurate and reduces
821 if (afi_clocks <= 0x100) {
822 writel(SKIP_DELAY_LOOP_VALUE_OR_ZERO(inner),
823 &sdr_rw_load_mgr_regs->load_cntr1);
825 writel(RW_MGR_IDLE_LOOP1,
826 &sdr_rw_load_jump_mgr_regs->load_jump_add1);
828 writel(RW_MGR_IDLE_LOOP1, SDR_PHYGRP_RWMGRGRP_ADDRESS |
829 RW_MGR_RUN_SINGLE_GROUP_OFFSET);
831 writel(SKIP_DELAY_LOOP_VALUE_OR_ZERO(inner),
832 &sdr_rw_load_mgr_regs->load_cntr0);
834 writel(SKIP_DELAY_LOOP_VALUE_OR_ZERO(outer),
835 &sdr_rw_load_mgr_regs->load_cntr1);
837 writel(RW_MGR_IDLE_LOOP2,
838 &sdr_rw_load_jump_mgr_regs->load_jump_add0);
840 writel(RW_MGR_IDLE_LOOP2,
841 &sdr_rw_load_jump_mgr_regs->load_jump_add1);
843 /* hack to get around compiler not being smart enough */
844 if (afi_clocks <= 0x10000) {
845 /* only need to run once */
846 writel(RW_MGR_IDLE_LOOP2, SDR_PHYGRP_RWMGRGRP_ADDRESS |
847 RW_MGR_RUN_SINGLE_GROUP_OFFSET);
850 writel(RW_MGR_IDLE_LOOP2,
851 SDR_PHYGRP_RWMGRGRP_ADDRESS |
852 RW_MGR_RUN_SINGLE_GROUP_OFFSET);
853 } while (c_loop-- != 0);
856 debug("%s:%d clocks=%u ... end\n", __func__, __LINE__, clocks);
859 static void rw_mgr_mem_initialize(void)
862 uint32_t grpaddr = SDR_PHYGRP_RWMGRGRP_ADDRESS |
863 RW_MGR_RUN_SINGLE_GROUP_OFFSET;
865 debug("%s:%d\n", __func__, __LINE__);
867 /* The reset / cke part of initialization is broadcasted to all ranks */
868 writel(RW_MGR_RANK_ALL, SDR_PHYGRP_RWMGRGRP_ADDRESS |
869 RW_MGR_SET_CS_AND_ODT_MASK_OFFSET);
872 * Here's how you load register for a loop
873 * Counters are located @ 0x800
874 * Jump address are located @ 0xC00
875 * For both, registers 0 to 3 are selected using bits 3 and 2, like
876 * in 0x800, 0x804, 0x808, 0x80C and 0xC00, 0xC04, 0xC08, 0xC0C
877 * I know this ain't pretty, but Avalon bus throws away the 2 least
881 /* start with memory RESET activated */
886 * 200us @ 266MHz (3.75 ns) ~ 54000 clock cycles
887 * If a and b are the number of iteration in 2 nested loops
888 * it takes the following number of cycles to complete the operation:
889 * number_of_cycles = ((2 + n) * a + 2) * b
890 * where n is the number of instruction in the inner loop
891 * One possible solution is n = 0 , a = 256 , b = 106 => a = FF,
896 writel(SKIP_DELAY_LOOP_VALUE_OR_ZERO(SEQ_TINIT_CNTR0_VAL),
897 &sdr_rw_load_mgr_regs->load_cntr0);
898 writel(SKIP_DELAY_LOOP_VALUE_OR_ZERO(SEQ_TINIT_CNTR1_VAL),
899 &sdr_rw_load_mgr_regs->load_cntr1);
900 writel(SKIP_DELAY_LOOP_VALUE_OR_ZERO(SEQ_TINIT_CNTR2_VAL),
901 &sdr_rw_load_mgr_regs->load_cntr2);
903 /* Load jump address */
904 writel(RW_MGR_INIT_RESET_0_CKE_0,
905 &sdr_rw_load_jump_mgr_regs->load_jump_add0);
906 writel(RW_MGR_INIT_RESET_0_CKE_0,
907 &sdr_rw_load_jump_mgr_regs->load_jump_add1);
908 writel(RW_MGR_INIT_RESET_0_CKE_0,
909 &sdr_rw_load_jump_mgr_regs->load_jump_add2);
911 /* Execute count instruction */
912 writel(RW_MGR_INIT_RESET_0_CKE_0, grpaddr);
914 /* indicate that memory is stable */
915 writel(1, &phy_mgr_cfg->reset_mem_stbl);
918 * transition the RESET to high
923 * 500us @ 266MHz (3.75 ns) ~ 134000 clock cycles
924 * If a and b are the number of iteration in 2 nested loops
925 * it takes the following number of cycles to complete the operation
926 * number_of_cycles = ((2 + n) * a + 2) * b
927 * where n is the number of instruction in the inner loop
928 * One possible solution is n = 2 , a = 131 , b = 256 => a = 83,
933 writel(SKIP_DELAY_LOOP_VALUE_OR_ZERO(SEQ_TRESET_CNTR0_VAL),
934 &sdr_rw_load_mgr_regs->load_cntr0);
935 writel(SKIP_DELAY_LOOP_VALUE_OR_ZERO(SEQ_TRESET_CNTR1_VAL),
936 &sdr_rw_load_mgr_regs->load_cntr1);
937 writel(SKIP_DELAY_LOOP_VALUE_OR_ZERO(SEQ_TRESET_CNTR2_VAL),
938 &sdr_rw_load_mgr_regs->load_cntr2);
940 /* Load jump address */
941 writel(RW_MGR_INIT_RESET_1_CKE_0,
942 &sdr_rw_load_jump_mgr_regs->load_jump_add0);
943 writel(RW_MGR_INIT_RESET_1_CKE_0,
944 &sdr_rw_load_jump_mgr_regs->load_jump_add1);
945 writel(RW_MGR_INIT_RESET_1_CKE_0,
946 &sdr_rw_load_jump_mgr_regs->load_jump_add2);
948 writel(RW_MGR_INIT_RESET_1_CKE_0, grpaddr);
950 /* bring up clock enable */
952 /* tXRP < 250 ck cycles */
953 delay_for_n_mem_clocks(250);
955 for (r = 0; r < RW_MGR_MEM_NUMBER_OF_RANKS; r++) {
956 if (param->skip_ranks[r]) {
957 /* request to skip the rank */
962 set_rank_and_odt_mask(r, RW_MGR_ODT_MODE_OFF);
965 * USER Use Mirror-ed commands for odd ranks if address
968 if ((RW_MGR_MEM_ADDRESS_MIRRORING >> r) & 0x1) {
969 set_jump_as_return();
970 writel(RW_MGR_MRS2_MIRR, grpaddr);
971 delay_for_n_mem_clocks(4);
972 set_jump_as_return();
973 writel(RW_MGR_MRS3_MIRR, grpaddr);
974 delay_for_n_mem_clocks(4);
975 set_jump_as_return();
976 writel(RW_MGR_MRS1_MIRR, grpaddr);
977 delay_for_n_mem_clocks(4);
978 set_jump_as_return();
979 writel(RW_MGR_MRS0_DLL_RESET_MIRR, grpaddr);
981 set_jump_as_return();
982 writel(RW_MGR_MRS2, grpaddr);
983 delay_for_n_mem_clocks(4);
984 set_jump_as_return();
985 writel(RW_MGR_MRS3, grpaddr);
986 delay_for_n_mem_clocks(4);
987 set_jump_as_return();
988 writel(RW_MGR_MRS1, grpaddr);
989 set_jump_as_return();
990 writel(RW_MGR_MRS0_DLL_RESET, grpaddr);
992 set_jump_as_return();
993 writel(RW_MGR_ZQCL, grpaddr);
995 /* tZQinit = tDLLK = 512 ck cycles */
996 delay_for_n_mem_clocks(512);
1001 * At the end of calibration we have to program the user settings in, and
1002 * USER hand off the memory to the user.
1004 static void rw_mgr_mem_handoff(void)
1007 uint32_t grpaddr = SDR_PHYGRP_RWMGRGRP_ADDRESS |
1008 RW_MGR_RUN_SINGLE_GROUP_OFFSET;
1010 debug("%s:%d\n", __func__, __LINE__);
1011 for (r = 0; r < RW_MGR_MEM_NUMBER_OF_RANKS; r++) {
1012 if (param->skip_ranks[r])
1013 /* request to skip the rank */
1016 set_rank_and_odt_mask(r, RW_MGR_ODT_MODE_OFF);
1018 /* precharge all banks ... */
1019 writel(RW_MGR_PRECHARGE_ALL, grpaddr);
1021 /* load up MR settings specified by user */
1024 * Use Mirror-ed commands for odd ranks if address
1027 if ((RW_MGR_MEM_ADDRESS_MIRRORING >> r) & 0x1) {
1028 set_jump_as_return();
1029 writel(RW_MGR_MRS2_MIRR, grpaddr);
1030 delay_for_n_mem_clocks(4);
1031 set_jump_as_return();
1032 writel(RW_MGR_MRS3_MIRR, grpaddr);
1033 delay_for_n_mem_clocks(4);
1034 set_jump_as_return();
1035 writel(RW_MGR_MRS1_MIRR, grpaddr);
1036 delay_for_n_mem_clocks(4);
1037 set_jump_as_return();
1038 writel(RW_MGR_MRS0_USER_MIRR, grpaddr);
1040 set_jump_as_return();
1041 writel(RW_MGR_MRS2, grpaddr);
1042 delay_for_n_mem_clocks(4);
1043 set_jump_as_return();
1044 writel(RW_MGR_MRS3, grpaddr);
1045 delay_for_n_mem_clocks(4);
1046 set_jump_as_return();
1047 writel(RW_MGR_MRS1, grpaddr);
1048 delay_for_n_mem_clocks(4);
1049 set_jump_as_return();
1050 writel(RW_MGR_MRS0_USER, grpaddr);
1053 * USER need to wait tMOD (12CK or 15ns) time before issuing
1054 * other commands, but we will have plenty of NIOS cycles before
1055 * actual handoff so its okay.
1061 * performs a guaranteed read on the patterns we are going to use during a
1062 * read test to ensure memory works
1064 static uint32_t rw_mgr_mem_calibrate_read_test_patterns(uint32_t rank_bgn,
1065 uint32_t group, uint32_t num_tries, uint32_t *bit_chk,
1069 uint32_t correct_mask_vg;
1070 uint32_t tmp_bit_chk;
1071 uint32_t rank_end = all_ranks ? RW_MGR_MEM_NUMBER_OF_RANKS :
1072 (rank_bgn + NUM_RANKS_PER_SHADOW_REG);
1074 uint32_t base_rw_mgr;
1076 *bit_chk = param->read_correct_mask;
1077 correct_mask_vg = param->read_correct_mask_vg;
1079 for (r = rank_bgn; r < rank_end; r++) {
1080 if (param->skip_ranks[r])
1081 /* request to skip the rank */
1085 set_rank_and_odt_mask(r, RW_MGR_ODT_MODE_READ_WRITE);
1087 /* Load up a constant bursts of read commands */
1088 writel(0x20, &sdr_rw_load_mgr_regs->load_cntr0);
1089 writel(RW_MGR_GUARANTEED_READ,
1090 &sdr_rw_load_jump_mgr_regs->load_jump_add0);
1092 writel(0x20, &sdr_rw_load_mgr_regs->load_cntr1);
1093 writel(RW_MGR_GUARANTEED_READ_CONT,
1094 &sdr_rw_load_jump_mgr_regs->load_jump_add1);
1097 for (vg = RW_MGR_MEM_VIRTUAL_GROUPS_PER_READ_DQS-1; ; vg--) {
1098 /* reset the fifos to get pointers to known state */
1100 writel(0, &phy_mgr_cmd->fifo_reset);
1101 writel(0, SDR_PHYGRP_RWMGRGRP_ADDRESS |
1102 RW_MGR_RESET_READ_DATAPATH_OFFSET);
1104 tmp_bit_chk = tmp_bit_chk << (RW_MGR_MEM_DQ_PER_READ_DQS
1105 / RW_MGR_MEM_VIRTUAL_GROUPS_PER_READ_DQS);
1107 addr = SDR_PHYGRP_RWMGRGRP_ADDRESS | RW_MGR_RUN_SINGLE_GROUP_OFFSET;
1108 writel(RW_MGR_GUARANTEED_READ, addr +
1109 ((group * RW_MGR_MEM_VIRTUAL_GROUPS_PER_READ_DQS +
1112 base_rw_mgr = readl(SDR_PHYGRP_RWMGRGRP_ADDRESS);
1113 tmp_bit_chk = tmp_bit_chk | (correct_mask_vg & (~base_rw_mgr));
1118 *bit_chk &= tmp_bit_chk;
1121 addr = SDR_PHYGRP_RWMGRGRP_ADDRESS | RW_MGR_RUN_SINGLE_GROUP_OFFSET;
1122 writel(RW_MGR_CLEAR_DQS_ENABLE, addr + (group << 2));
1124 set_rank_and_odt_mask(0, RW_MGR_ODT_MODE_OFF);
1125 debug_cond(DLEVEL == 1, "%s:%d test_load_patterns(%u,ALL) => (%u == %u) =>\
1126 %lu\n", __func__, __LINE__, group, *bit_chk, param->read_correct_mask,
1127 (long unsigned int)(*bit_chk == param->read_correct_mask));
1128 return *bit_chk == param->read_correct_mask;
1131 static uint32_t rw_mgr_mem_calibrate_read_test_patterns_all_ranks
1132 (uint32_t group, uint32_t num_tries, uint32_t *bit_chk)
1134 return rw_mgr_mem_calibrate_read_test_patterns(0, group,
1135 num_tries, bit_chk, 1);
1138 /* load up the patterns we are going to use during a read test */
1139 static void rw_mgr_mem_calibrate_read_load_patterns(uint32_t rank_bgn,
1143 uint32_t rank_end = all_ranks ? RW_MGR_MEM_NUMBER_OF_RANKS :
1144 (rank_bgn + NUM_RANKS_PER_SHADOW_REG);
1146 debug("%s:%d\n", __func__, __LINE__);
1147 for (r = rank_bgn; r < rank_end; r++) {
1148 if (param->skip_ranks[r])
1149 /* request to skip the rank */
1153 set_rank_and_odt_mask(r, RW_MGR_ODT_MODE_READ_WRITE);
1155 /* Load up a constant bursts */
1156 writel(0x20, &sdr_rw_load_mgr_regs->load_cntr0);
1158 writel(RW_MGR_GUARANTEED_WRITE_WAIT0,
1159 &sdr_rw_load_jump_mgr_regs->load_jump_add0);
1161 writel(0x20, &sdr_rw_load_mgr_regs->load_cntr1);
1163 writel(RW_MGR_GUARANTEED_WRITE_WAIT1,
1164 &sdr_rw_load_jump_mgr_regs->load_jump_add1);
1166 writel(0x04, &sdr_rw_load_mgr_regs->load_cntr2);
1168 writel(RW_MGR_GUARANTEED_WRITE_WAIT2,
1169 &sdr_rw_load_jump_mgr_regs->load_jump_add2);
1171 writel(0x04, &sdr_rw_load_mgr_regs->load_cntr3);
1173 writel(RW_MGR_GUARANTEED_WRITE_WAIT3,
1174 &sdr_rw_load_jump_mgr_regs->load_jump_add3);
1176 writel(RW_MGR_GUARANTEED_WRITE, SDR_PHYGRP_RWMGRGRP_ADDRESS |
1177 RW_MGR_RUN_SINGLE_GROUP_OFFSET);
1180 set_rank_and_odt_mask(0, RW_MGR_ODT_MODE_OFF);
1184 * try a read and see if it returns correct data back. has dummy reads
1185 * inserted into the mix used to align dqs enable. has more thorough checks
1186 * than the regular read test.
1188 static uint32_t rw_mgr_mem_calibrate_read_test(uint32_t rank_bgn, uint32_t group,
1189 uint32_t num_tries, uint32_t all_correct, uint32_t *bit_chk,
1190 uint32_t all_groups, uint32_t all_ranks)
1193 uint32_t correct_mask_vg;
1194 uint32_t tmp_bit_chk;
1195 uint32_t rank_end = all_ranks ? RW_MGR_MEM_NUMBER_OF_RANKS :
1196 (rank_bgn + NUM_RANKS_PER_SHADOW_REG);
1198 uint32_t base_rw_mgr;
1200 *bit_chk = param->read_correct_mask;
1201 correct_mask_vg = param->read_correct_mask_vg;
1203 uint32_t quick_read_mode = (((STATIC_CALIB_STEPS) &
1204 CALIB_SKIP_DELAY_SWEEPS) && ENABLE_SUPER_QUICK_CALIBRATION);
1206 for (r = rank_bgn; r < rank_end; r++) {
1207 if (param->skip_ranks[r])
1208 /* request to skip the rank */
1212 set_rank_and_odt_mask(r, RW_MGR_ODT_MODE_READ_WRITE);
1214 writel(0x10, &sdr_rw_load_mgr_regs->load_cntr1);
1216 writel(RW_MGR_READ_B2B_WAIT1,
1217 &sdr_rw_load_jump_mgr_regs->load_jump_add1);
1219 writel(0x10, &sdr_rw_load_mgr_regs->load_cntr2);
1220 writel(RW_MGR_READ_B2B_WAIT2,
1221 &sdr_rw_load_jump_mgr_regs->load_jump_add2);
1223 if (quick_read_mode)
1224 writel(0x1, &sdr_rw_load_mgr_regs->load_cntr0);
1225 /* need at least two (1+1) reads to capture failures */
1226 else if (all_groups)
1227 writel(0x06, &sdr_rw_load_mgr_regs->load_cntr0);
1229 writel(0x32, &sdr_rw_load_mgr_regs->load_cntr0);
1231 writel(RW_MGR_READ_B2B,
1232 &sdr_rw_load_jump_mgr_regs->load_jump_add0);
1234 writel(RW_MGR_MEM_IF_READ_DQS_WIDTH *
1235 RW_MGR_MEM_VIRTUAL_GROUPS_PER_READ_DQS - 1,
1236 &sdr_rw_load_mgr_regs->load_cntr3);
1238 writel(0x0, &sdr_rw_load_mgr_regs->load_cntr3);
1240 writel(RW_MGR_READ_B2B,
1241 &sdr_rw_load_jump_mgr_regs->load_jump_add3);
1244 for (vg = RW_MGR_MEM_VIRTUAL_GROUPS_PER_READ_DQS-1; ; vg--) {
1245 /* reset the fifos to get pointers to known state */
1246 writel(0, &phy_mgr_cmd->fifo_reset);
1247 writel(0, SDR_PHYGRP_RWMGRGRP_ADDRESS |
1248 RW_MGR_RESET_READ_DATAPATH_OFFSET);
1250 tmp_bit_chk = tmp_bit_chk << (RW_MGR_MEM_DQ_PER_READ_DQS
1251 / RW_MGR_MEM_VIRTUAL_GROUPS_PER_READ_DQS);
1254 addr = SDR_PHYGRP_RWMGRGRP_ADDRESS | RW_MGR_RUN_ALL_GROUPS_OFFSET;
1256 addr = SDR_PHYGRP_RWMGRGRP_ADDRESS | RW_MGR_RUN_SINGLE_GROUP_OFFSET;
1258 writel(RW_MGR_READ_B2B, addr +
1259 ((group * RW_MGR_MEM_VIRTUAL_GROUPS_PER_READ_DQS +
1262 base_rw_mgr = readl(SDR_PHYGRP_RWMGRGRP_ADDRESS);
1263 tmp_bit_chk = tmp_bit_chk | (correct_mask_vg & ~(base_rw_mgr));
1268 *bit_chk &= tmp_bit_chk;
1271 addr = SDR_PHYGRP_RWMGRGRP_ADDRESS | RW_MGR_RUN_SINGLE_GROUP_OFFSET;
1272 writel(RW_MGR_CLEAR_DQS_ENABLE, addr + (group << 2));
1275 set_rank_and_odt_mask(0, RW_MGR_ODT_MODE_OFF);
1276 debug_cond(DLEVEL == 2, "%s:%d read_test(%u,ALL,%u) =>\
1277 (%u == %u) => %lu", __func__, __LINE__, group,
1278 all_groups, *bit_chk, param->read_correct_mask,
1279 (long unsigned int)(*bit_chk ==
1280 param->read_correct_mask));
1281 return *bit_chk == param->read_correct_mask;
1283 set_rank_and_odt_mask(0, RW_MGR_ODT_MODE_OFF);
1284 debug_cond(DLEVEL == 2, "%s:%d read_test(%u,ONE,%u) =>\
1285 (%u != %lu) => %lu\n", __func__, __LINE__,
1286 group, all_groups, *bit_chk, (long unsigned int)0,
1287 (long unsigned int)(*bit_chk != 0x00));
1288 return *bit_chk != 0x00;
1292 static uint32_t rw_mgr_mem_calibrate_read_test_all_ranks(uint32_t group,
1293 uint32_t num_tries, uint32_t all_correct, uint32_t *bit_chk,
1294 uint32_t all_groups)
1296 return rw_mgr_mem_calibrate_read_test(0, group, num_tries, all_correct,
1297 bit_chk, all_groups, 1);
1300 static void rw_mgr_incr_vfifo(uint32_t grp, uint32_t *v)
1302 writel(grp, &phy_mgr_cmd->inc_vfifo_hard_phy);
1306 static void rw_mgr_decr_vfifo(uint32_t grp, uint32_t *v)
1310 for (i = 0; i < VFIFO_SIZE-1; i++)
1311 rw_mgr_incr_vfifo(grp, v);
1314 static int find_vfifo_read(uint32_t grp, uint32_t *bit_chk)
1317 uint32_t fail_cnt = 0;
1318 uint32_t test_status;
1320 for (v = 0; v < VFIFO_SIZE; ) {
1321 debug_cond(DLEVEL == 2, "%s:%d find_dqs_en_phase: vfifo %u\n",
1322 __func__, __LINE__, v);
1323 test_status = rw_mgr_mem_calibrate_read_test_all_ranks
1324 (grp, 1, PASS_ONE_BIT, bit_chk, 0);
1332 /* fiddle with FIFO */
1333 rw_mgr_incr_vfifo(grp, &v);
1336 if (v >= VFIFO_SIZE) {
1337 /* no failing read found!! Something must have gone wrong */
1338 debug_cond(DLEVEL == 2, "%s:%d find_dqs_en_phase: vfifo failed\n",
1339 __func__, __LINE__);
1346 static int find_working_phase(uint32_t *grp, uint32_t *bit_chk,
1347 uint32_t dtaps_per_ptap, uint32_t *work_bgn,
1348 uint32_t *v, uint32_t *d, uint32_t *p,
1349 uint32_t *i, uint32_t *max_working_cnt)
1351 uint32_t found_begin = 0;
1352 uint32_t tmp_delay = 0;
1353 uint32_t test_status;
1355 for (*d = 0; *d <= dtaps_per_ptap; (*d)++, tmp_delay +=
1356 IO_DELAY_PER_DQS_EN_DCHAIN_TAP) {
1357 *work_bgn = tmp_delay;
1358 scc_mgr_set_dqs_en_delay_all_ranks(*grp, *d);
1360 for (*i = 0; *i < VFIFO_SIZE; (*i)++) {
1361 for (*p = 0; *p <= IO_DQS_EN_PHASE_MAX; (*p)++, *work_bgn +=
1362 IO_DELAY_PER_OPA_TAP) {
1363 scc_mgr_set_dqs_en_phase_all_ranks(*grp, *p);
1366 rw_mgr_mem_calibrate_read_test_all_ranks
1367 (*grp, 1, PASS_ONE_BIT, bit_chk, 0);
1370 *max_working_cnt = 1;
1379 if (*p > IO_DQS_EN_PHASE_MAX)
1380 /* fiddle with FIFO */
1381 rw_mgr_incr_vfifo(*grp, v);
1388 if (*i >= VFIFO_SIZE) {
1389 /* cannot find working solution */
1390 debug_cond(DLEVEL == 2, "%s:%d find_dqs_en_phase: no vfifo/\
1391 ptap/dtap\n", __func__, __LINE__);
1398 static void sdr_backup_phase(uint32_t *grp, uint32_t *bit_chk,
1399 uint32_t *work_bgn, uint32_t *v, uint32_t *d,
1400 uint32_t *p, uint32_t *max_working_cnt)
1402 uint32_t found_begin = 0;
1405 /* Special case code for backing up a phase */
1407 *p = IO_DQS_EN_PHASE_MAX;
1408 rw_mgr_decr_vfifo(*grp, v);
1412 tmp_delay = *work_bgn - IO_DELAY_PER_OPA_TAP;
1413 scc_mgr_set_dqs_en_phase_all_ranks(*grp, *p);
1415 for (*d = 0; *d <= IO_DQS_EN_DELAY_MAX && tmp_delay < *work_bgn;
1416 (*d)++, tmp_delay += IO_DELAY_PER_DQS_EN_DCHAIN_TAP) {
1417 scc_mgr_set_dqs_en_delay_all_ranks(*grp, *d);
1419 if (rw_mgr_mem_calibrate_read_test_all_ranks(*grp, 1,
1423 *work_bgn = tmp_delay;
1428 /* We have found a working dtap before the ptap found above */
1429 if (found_begin == 1)
1430 (*max_working_cnt)++;
1433 * Restore VFIFO to old state before we decremented it
1437 if (*p > IO_DQS_EN_PHASE_MAX) {
1439 rw_mgr_incr_vfifo(*grp, v);
1442 scc_mgr_set_dqs_en_delay_all_ranks(*grp, 0);
1445 static int sdr_nonworking_phase(uint32_t *grp, uint32_t *bit_chk,
1446 uint32_t *work_bgn, uint32_t *v, uint32_t *d,
1447 uint32_t *p, uint32_t *i, uint32_t *max_working_cnt,
1450 uint32_t found_end = 0;
1453 *work_end += IO_DELAY_PER_OPA_TAP;
1454 if (*p > IO_DQS_EN_PHASE_MAX) {
1455 /* fiddle with FIFO */
1457 rw_mgr_incr_vfifo(*grp, v);
1460 for (; *i < VFIFO_SIZE + 1; (*i)++) {
1461 for (; *p <= IO_DQS_EN_PHASE_MAX; (*p)++, *work_end
1462 += IO_DELAY_PER_OPA_TAP) {
1463 scc_mgr_set_dqs_en_phase_all_ranks(*grp, *p);
1465 if (!rw_mgr_mem_calibrate_read_test_all_ranks
1466 (*grp, 1, PASS_ONE_BIT, bit_chk, 0)) {
1470 (*max_working_cnt)++;
1477 if (*p > IO_DQS_EN_PHASE_MAX) {
1478 /* fiddle with FIFO */
1479 rw_mgr_incr_vfifo(*grp, v);
1484 if (*i >= VFIFO_SIZE + 1) {
1485 /* cannot see edge of failing read */
1486 debug_cond(DLEVEL == 2, "%s:%d sdr_nonworking_phase: end:\
1487 failed\n", __func__, __LINE__);
1494 static int sdr_find_window_centre(uint32_t *grp, uint32_t *bit_chk,
1495 uint32_t *work_bgn, uint32_t *v, uint32_t *d,
1496 uint32_t *p, uint32_t *work_mid,
1502 *work_mid = (*work_bgn + *work_end) / 2;
1504 debug_cond(DLEVEL == 2, "work_bgn=%d work_end=%d work_mid=%d\n",
1505 *work_bgn, *work_end, *work_mid);
1506 /* Get the middle delay to be less than a VFIFO delay */
1507 for (*p = 0; *p <= IO_DQS_EN_PHASE_MAX;
1508 (*p)++, tmp_delay += IO_DELAY_PER_OPA_TAP)
1510 debug_cond(DLEVEL == 2, "vfifo ptap delay %d\n", tmp_delay);
1511 while (*work_mid > tmp_delay)
1512 *work_mid -= tmp_delay;
1513 debug_cond(DLEVEL == 2, "new work_mid %d\n", *work_mid);
1516 for (*p = 0; *p <= IO_DQS_EN_PHASE_MAX && tmp_delay < *work_mid;
1517 (*p)++, tmp_delay += IO_DELAY_PER_OPA_TAP)
1519 tmp_delay -= IO_DELAY_PER_OPA_TAP;
1520 debug_cond(DLEVEL == 2, "new p %d, tmp_delay=%d\n", (*p) - 1, tmp_delay);
1521 for (*d = 0; *d <= IO_DQS_EN_DELAY_MAX && tmp_delay < *work_mid; (*d)++,
1522 tmp_delay += IO_DELAY_PER_DQS_EN_DCHAIN_TAP)
1524 debug_cond(DLEVEL == 2, "new d %d, tmp_delay=%d\n", *d, tmp_delay);
1526 scc_mgr_set_dqs_en_phase_all_ranks(*grp, (*p) - 1);
1527 scc_mgr_set_dqs_en_delay_all_ranks(*grp, *d);
1530 * push vfifo until we can successfully calibrate. We can do this
1531 * because the largest possible margin in 1 VFIFO cycle.
1533 for (i = 0; i < VFIFO_SIZE; i++) {
1534 debug_cond(DLEVEL == 2, "find_dqs_en_phase: center: vfifo=%u\n",
1536 if (rw_mgr_mem_calibrate_read_test_all_ranks(*grp, 1,
1542 /* fiddle with FIFO */
1543 rw_mgr_incr_vfifo(*grp, v);
1546 if (i >= VFIFO_SIZE) {
1547 debug_cond(DLEVEL == 2, "%s:%d find_dqs_en_phase: center: \
1548 failed\n", __func__, __LINE__);
1555 /* find a good dqs enable to use */
1556 static uint32_t rw_mgr_mem_calibrate_vfifo_find_dqs_en_phase(uint32_t grp)
1558 uint32_t v, d, p, i;
1559 uint32_t max_working_cnt;
1561 uint32_t dtaps_per_ptap;
1562 uint32_t work_bgn, work_mid, work_end;
1563 uint32_t found_passing_read, found_failing_read, initial_failing_dtap;
1565 debug("%s:%d %u\n", __func__, __LINE__, grp);
1567 reg_file_set_sub_stage(CAL_SUBSTAGE_VFIFO_CENTER);
1569 scc_mgr_set_dqs_en_delay_all_ranks(grp, 0);
1570 scc_mgr_set_dqs_en_phase_all_ranks(grp, 0);
1572 /* ************************************************************** */
1573 /* * Step 0 : Determine number of delay taps for each phase tap * */
1574 dtaps_per_ptap = IO_DELAY_PER_OPA_TAP/IO_DELAY_PER_DQS_EN_DCHAIN_TAP;
1576 /* ********************************************************* */
1577 /* * Step 1 : First push vfifo until we get a failing read * */
1578 v = find_vfifo_read(grp, &bit_chk);
1580 max_working_cnt = 0;
1582 /* ******************************************************** */
1583 /* * step 2: find first working phase, increment in ptaps * */
1585 if (find_working_phase(&grp, &bit_chk, dtaps_per_ptap, &work_bgn, &v, &d,
1586 &p, &i, &max_working_cnt) == 0)
1589 work_end = work_bgn;
1592 * If d is 0 then the working window covers a phase tap and
1593 * we can follow the old procedure otherwise, we've found the beginning,
1594 * and we need to increment the dtaps until we find the end.
1597 /* ********************************************************* */
1598 /* * step 3a: if we have room, back off by one and
1599 increment in dtaps * */
1601 sdr_backup_phase(&grp, &bit_chk, &work_bgn, &v, &d, &p,
1604 /* ********************************************************* */
1605 /* * step 4a: go forward from working phase to non working
1606 phase, increment in ptaps * */
1607 if (sdr_nonworking_phase(&grp, &bit_chk, &work_bgn, &v, &d, &p,
1608 &i, &max_working_cnt, &work_end) == 0)
1611 /* ********************************************************* */
1612 /* * step 5a: back off one from last, increment in dtaps * */
1614 /* Special case code for backing up a phase */
1616 p = IO_DQS_EN_PHASE_MAX;
1617 rw_mgr_decr_vfifo(grp, &v);
1622 work_end -= IO_DELAY_PER_OPA_TAP;
1623 scc_mgr_set_dqs_en_phase_all_ranks(grp, p);
1625 /* * The actual increment of dtaps is done outside of
1626 the if/else loop to share code */
1629 debug_cond(DLEVEL == 2, "%s:%d find_dqs_en_phase: v/p: \
1630 vfifo=%u ptap=%u\n", __func__, __LINE__,
1633 /* ******************************************************* */
1634 /* * step 3-5b: Find the right edge of the window using
1636 debug_cond(DLEVEL == 2, "%s:%d find_dqs_en_phase:vfifo=%u \
1637 ptap=%u dtap=%u bgn=%u\n", __func__, __LINE__,
1640 work_end = work_bgn;
1642 /* * The actual increment of dtaps is done outside of the
1643 if/else loop to share code */
1645 /* Only here to counterbalance a subtract later on which is
1646 not needed if this branch of the algorithm is taken */
1650 /* The dtap increment to find the failing edge is done here */
1651 for (; d <= IO_DQS_EN_DELAY_MAX; d++, work_end +=
1652 IO_DELAY_PER_DQS_EN_DCHAIN_TAP) {
1653 debug_cond(DLEVEL == 2, "%s:%d find_dqs_en_phase: \
1654 end-2: dtap=%u\n", __func__, __LINE__, d);
1655 scc_mgr_set_dqs_en_delay_all_ranks(grp, d);
1657 if (!rw_mgr_mem_calibrate_read_test_all_ranks(grp, 1,
1664 /* Go back to working dtap */
1666 work_end -= IO_DELAY_PER_DQS_EN_DCHAIN_TAP;
1668 debug_cond(DLEVEL == 2, "%s:%d find_dqs_en_phase: v/p/d: vfifo=%u \
1669 ptap=%u dtap=%u end=%u\n", __func__, __LINE__,
1670 v, p, d-1, work_end);
1672 if (work_end < work_bgn) {
1674 debug_cond(DLEVEL == 2, "%s:%d find_dqs_en_phase: end-2: \
1675 failed\n", __func__, __LINE__);
1679 debug_cond(DLEVEL == 2, "%s:%d find_dqs_en_phase: found range [%u,%u]\n",
1680 __func__, __LINE__, work_bgn, work_end);
1682 /* *************************************************************** */
1684 * * We need to calculate the number of dtaps that equal a ptap
1685 * * To do that we'll back up a ptap and re-find the edge of the
1686 * * window using dtaps
1689 debug_cond(DLEVEL == 2, "%s:%d find_dqs_en_phase: calculate dtaps_per_ptap \
1690 for tracking\n", __func__, __LINE__);
1692 /* Special case code for backing up a phase */
1694 p = IO_DQS_EN_PHASE_MAX;
1695 rw_mgr_decr_vfifo(grp, &v);
1696 debug_cond(DLEVEL == 2, "%s:%d find_dqs_en_phase: backedup \
1697 cycle/phase: v=%u p=%u\n", __func__, __LINE__,
1701 debug_cond(DLEVEL == 2, "%s:%d find_dqs_en_phase: backedup \
1702 phase only: v=%u p=%u", __func__, __LINE__,
1706 scc_mgr_set_dqs_en_phase_all_ranks(grp, p);
1709 * Increase dtap until we first see a passing read (in case the
1710 * window is smaller than a ptap),
1711 * and then a failing read to mark the edge of the window again
1714 /* Find a passing read */
1715 debug_cond(DLEVEL == 2, "%s:%d find_dqs_en_phase: find passing read\n",
1716 __func__, __LINE__);
1717 found_passing_read = 0;
1718 found_failing_read = 0;
1719 initial_failing_dtap = d;
1720 for (; d <= IO_DQS_EN_DELAY_MAX; d++) {
1721 debug_cond(DLEVEL == 2, "%s:%d find_dqs_en_phase: testing \
1722 read d=%u\n", __func__, __LINE__, d);
1723 scc_mgr_set_dqs_en_delay_all_ranks(grp, d);
1725 if (rw_mgr_mem_calibrate_read_test_all_ranks(grp, 1,
1728 found_passing_read = 1;
1733 if (found_passing_read) {
1734 /* Find a failing read */
1735 debug_cond(DLEVEL == 2, "%s:%d find_dqs_en_phase: find failing \
1736 read\n", __func__, __LINE__);
1737 for (d = d + 1; d <= IO_DQS_EN_DELAY_MAX; d++) {
1738 debug_cond(DLEVEL == 2, "%s:%d find_dqs_en_phase: \
1739 testing read d=%u\n", __func__, __LINE__, d);
1740 scc_mgr_set_dqs_en_delay_all_ranks(grp, d);
1742 if (!rw_mgr_mem_calibrate_read_test_all_ranks
1743 (grp, 1, PASS_ONE_BIT, &bit_chk, 0)) {
1744 found_failing_read = 1;
1749 debug_cond(DLEVEL == 1, "%s:%d find_dqs_en_phase: failed to \
1750 calculate dtaps", __func__, __LINE__);
1751 debug_cond(DLEVEL == 1, "per ptap. Fall back on static value\n");
1755 * The dynamically calculated dtaps_per_ptap is only valid if we
1756 * found a passing/failing read. If we didn't, it means d hit the max
1757 * (IO_DQS_EN_DELAY_MAX). Otherwise, dtaps_per_ptap retains its
1758 * statically calculated value.
1760 if (found_passing_read && found_failing_read)
1761 dtaps_per_ptap = d - initial_failing_dtap;
1763 writel(dtaps_per_ptap, &sdr_reg_file->dtaps_per_ptap);
1764 debug_cond(DLEVEL == 2, "%s:%d find_dqs_en_phase: dtaps_per_ptap=%u \
1765 - %u = %u", __func__, __LINE__, d,
1766 initial_failing_dtap, dtaps_per_ptap);
1768 /* ******************************************** */
1769 /* * step 6: Find the centre of the window * */
1770 if (sdr_find_window_centre(&grp, &bit_chk, &work_bgn, &v, &d, &p,
1771 &work_mid, &work_end) == 0)
1774 debug_cond(DLEVEL == 2, "%s:%d find_dqs_en_phase: center found: \
1775 vfifo=%u ptap=%u dtap=%u\n", __func__, __LINE__,
1781 * Try rw_mgr_mem_calibrate_vfifo_find_dqs_en_phase across different
1782 * dq_in_delay values
1785 rw_mgr_mem_calibrate_vfifo_find_dqs_en_phase_sweep_dq_in_delay
1786 (uint32_t write_group, uint32_t read_group, uint32_t test_bgn)
1794 const uint32_t delay_step = IO_IO_IN_DELAY_MAX /
1795 (RW_MGR_MEM_DQ_PER_READ_DQS-1);
1796 /* we start at zero, so have one less dq to devide among */
1798 debug("%s:%d (%u,%u,%u)", __func__, __LINE__, write_group, read_group,
1801 /* try different dq_in_delays since the dq path is shorter than dqs */
1803 for (r = 0; r < RW_MGR_MEM_NUMBER_OF_RANKS;
1804 r += NUM_RANKS_PER_SHADOW_REG) {
1805 for (i = 0, p = test_bgn, d = 0; i < RW_MGR_MEM_DQ_PER_READ_DQS;
1806 i++, p++, d += delay_step) {
1807 debug_cond(DLEVEL == 1, "%s:%d rw_mgr_mem_calibrate_\
1808 vfifo_find_dqs_", __func__, __LINE__);
1809 debug_cond(DLEVEL == 1, "en_phase_sweep_dq_in_delay: g=%u/%u ",
1810 write_group, read_group);
1811 debug_cond(DLEVEL == 1, "r=%u, i=%u p=%u d=%u\n", r, i , p, d);
1812 scc_mgr_set_dq_in_delay(p, d);
1815 writel(0, &sdr_scc_mgr->update);
1818 found = rw_mgr_mem_calibrate_vfifo_find_dqs_en_phase(read_group);
1820 debug_cond(DLEVEL == 1, "%s:%d rw_mgr_mem_calibrate_vfifo_find_dqs_\
1821 en_phase_sweep_dq", __func__, __LINE__);
1822 debug_cond(DLEVEL == 1, "_in_delay: g=%u/%u found=%u; Reseting delay \
1823 chain to zero\n", write_group, read_group, found);
1825 for (r = 0; r < RW_MGR_MEM_NUMBER_OF_RANKS;
1826 r += NUM_RANKS_PER_SHADOW_REG) {
1827 for (i = 0, p = test_bgn; i < RW_MGR_MEM_DQ_PER_READ_DQS;
1829 scc_mgr_set_dq_in_delay(p, 0);
1832 writel(0, &sdr_scc_mgr->update);
1838 /* per-bit deskew DQ and center */
1839 static uint32_t rw_mgr_mem_calibrate_vfifo_center(uint32_t rank_bgn,
1840 uint32_t write_group, uint32_t read_group, uint32_t test_bgn,
1841 uint32_t use_read_test, uint32_t update_fom)
1843 uint32_t i, p, d, min_index;
1845 * Store these as signed since there are comparisons with
1849 uint32_t sticky_bit_chk;
1850 int32_t left_edge[RW_MGR_MEM_DQ_PER_READ_DQS];
1851 int32_t right_edge[RW_MGR_MEM_DQ_PER_READ_DQS];
1852 int32_t final_dq[RW_MGR_MEM_DQ_PER_READ_DQS];
1854 int32_t orig_mid_min, mid_min;
1855 int32_t new_dqs, start_dqs, start_dqs_en, shift_dq, final_dqs,
1857 int32_t dq_margin, dqs_margin;
1859 uint32_t temp_dq_in_delay1, temp_dq_in_delay2;
1862 debug("%s:%d: %u %u", __func__, __LINE__, read_group, test_bgn);
1864 addr = SDR_PHYGRP_SCCGRP_ADDRESS | SCC_MGR_DQS_IN_DELAY_OFFSET;
1865 start_dqs = readl(addr + (read_group << 2));
1866 if (IO_SHIFT_DQS_EN_WHEN_SHIFT_DQS)
1867 start_dqs_en = readl(addr + ((read_group << 2)
1868 - IO_DQS_EN_DELAY_OFFSET));
1870 /* set the left and right edge of each bit to an illegal value */
1871 /* use (IO_IO_IN_DELAY_MAX + 1) as an illegal value */
1873 for (i = 0; i < RW_MGR_MEM_DQ_PER_READ_DQS; i++) {
1874 left_edge[i] = IO_IO_IN_DELAY_MAX + 1;
1875 right_edge[i] = IO_IO_IN_DELAY_MAX + 1;
1878 /* Search for the left edge of the window for each bit */
1879 for (d = 0; d <= IO_IO_IN_DELAY_MAX; d++) {
1880 scc_mgr_apply_group_dq_in_delay(write_group, test_bgn, d);
1882 writel(0, &sdr_scc_mgr->update);
1885 * Stop searching when the read test doesn't pass AND when
1886 * we've seen a passing read on every bit.
1888 if (use_read_test) {
1889 stop = !rw_mgr_mem_calibrate_read_test(rank_bgn,
1890 read_group, NUM_READ_PB_TESTS, PASS_ONE_BIT,
1893 rw_mgr_mem_calibrate_write_test(rank_bgn, write_group,
1896 bit_chk = bit_chk >> (RW_MGR_MEM_DQ_PER_READ_DQS *
1897 (read_group - (write_group *
1898 RW_MGR_MEM_IF_READ_DQS_WIDTH /
1899 RW_MGR_MEM_IF_WRITE_DQS_WIDTH)));
1900 stop = (bit_chk == 0);
1902 sticky_bit_chk = sticky_bit_chk | bit_chk;
1903 stop = stop && (sticky_bit_chk == param->read_correct_mask);
1904 debug_cond(DLEVEL == 2, "%s:%d vfifo_center(left): dtap=%u => %u == %u \
1905 && %u", __func__, __LINE__, d,
1907 param->read_correct_mask, stop);
1912 for (i = 0; i < RW_MGR_MEM_DQ_PER_READ_DQS; i++) {
1914 /* Remember a passing test as the
1918 /* If a left edge has not been seen yet,
1919 then a future passing test will mark
1920 this edge as the right edge */
1922 IO_IO_IN_DELAY_MAX + 1) {
1923 right_edge[i] = -(d + 1);
1926 bit_chk = bit_chk >> 1;
1931 /* Reset DQ delay chains to 0 */
1932 scc_mgr_apply_group_dq_in_delay(write_group, test_bgn, 0);
1934 for (i = RW_MGR_MEM_DQ_PER_READ_DQS - 1;; i--) {
1935 debug_cond(DLEVEL == 2, "%s:%d vfifo_center: left_edge[%u]: \
1936 %d right_edge[%u]: %d\n", __func__, __LINE__,
1937 i, left_edge[i], i, right_edge[i]);
1940 * Check for cases where we haven't found the left edge,
1941 * which makes our assignment of the the right edge invalid.
1942 * Reset it to the illegal value.
1944 if ((left_edge[i] == IO_IO_IN_DELAY_MAX + 1) && (
1945 right_edge[i] != IO_IO_IN_DELAY_MAX + 1)) {
1946 right_edge[i] = IO_IO_IN_DELAY_MAX + 1;
1947 debug_cond(DLEVEL == 2, "%s:%d vfifo_center: reset \
1948 right_edge[%u]: %d\n", __func__, __LINE__,
1953 * Reset sticky bit (except for bits where we have seen
1954 * both the left and right edge).
1956 sticky_bit_chk = sticky_bit_chk << 1;
1957 if ((left_edge[i] != IO_IO_IN_DELAY_MAX + 1) &&
1958 (right_edge[i] != IO_IO_IN_DELAY_MAX + 1)) {
1959 sticky_bit_chk = sticky_bit_chk | 1;
1966 /* Search for the right edge of the window for each bit */
1967 for (d = 0; d <= IO_DQS_IN_DELAY_MAX - start_dqs; d++) {
1968 scc_mgr_set_dqs_bus_in_delay(read_group, d + start_dqs);
1969 if (IO_SHIFT_DQS_EN_WHEN_SHIFT_DQS) {
1970 uint32_t delay = d + start_dqs_en;
1971 if (delay > IO_DQS_EN_DELAY_MAX)
1972 delay = IO_DQS_EN_DELAY_MAX;
1973 scc_mgr_set_dqs_en_delay(read_group, delay);
1975 scc_mgr_load_dqs(read_group);
1977 writel(0, &sdr_scc_mgr->update);
1980 * Stop searching when the read test doesn't pass AND when
1981 * we've seen a passing read on every bit.
1983 if (use_read_test) {
1984 stop = !rw_mgr_mem_calibrate_read_test(rank_bgn,
1985 read_group, NUM_READ_PB_TESTS, PASS_ONE_BIT,
1988 rw_mgr_mem_calibrate_write_test(rank_bgn, write_group,
1991 bit_chk = bit_chk >> (RW_MGR_MEM_DQ_PER_READ_DQS *
1992 (read_group - (write_group *
1993 RW_MGR_MEM_IF_READ_DQS_WIDTH /
1994 RW_MGR_MEM_IF_WRITE_DQS_WIDTH)));
1995 stop = (bit_chk == 0);
1997 sticky_bit_chk = sticky_bit_chk | bit_chk;
1998 stop = stop && (sticky_bit_chk == param->read_correct_mask);
2000 debug_cond(DLEVEL == 2, "%s:%d vfifo_center(right): dtap=%u => %u == \
2001 %u && %u", __func__, __LINE__, d,
2002 sticky_bit_chk, param->read_correct_mask, stop);
2007 for (i = 0; i < RW_MGR_MEM_DQ_PER_READ_DQS; i++) {
2009 /* Remember a passing test as
2014 /* If a right edge has not been
2015 seen yet, then a future passing
2016 test will mark this edge as the
2018 if (right_edge[i] ==
2019 IO_IO_IN_DELAY_MAX + 1) {
2020 left_edge[i] = -(d + 1);
2023 /* d = 0 failed, but it passed
2024 when testing the left edge,
2025 so it must be marginal,
2027 if (right_edge[i] ==
2028 IO_IO_IN_DELAY_MAX + 1 &&
2034 /* If a right edge has not been
2035 seen yet, then a future passing
2036 test will mark this edge as the
2038 else if (right_edge[i] ==
2039 IO_IO_IN_DELAY_MAX +
2041 left_edge[i] = -(d + 1);
2046 debug_cond(DLEVEL == 2, "%s:%d vfifo_center[r,\
2047 d=%u]: ", __func__, __LINE__, d);
2048 debug_cond(DLEVEL == 2, "bit_chk_test=%d left_edge[%u]: %d ",
2049 (int)(bit_chk & 1), i, left_edge[i]);
2050 debug_cond(DLEVEL == 2, "right_edge[%u]: %d\n", i,
2052 bit_chk = bit_chk >> 1;
2057 /* Check that all bits have a window */
2058 for (i = 0; i < RW_MGR_MEM_DQ_PER_READ_DQS; i++) {
2059 debug_cond(DLEVEL == 2, "%s:%d vfifo_center: left_edge[%u]: \
2060 %d right_edge[%u]: %d", __func__, __LINE__,
2061 i, left_edge[i], i, right_edge[i]);
2062 if ((left_edge[i] == IO_IO_IN_DELAY_MAX + 1) || (right_edge[i]
2063 == IO_IO_IN_DELAY_MAX + 1)) {
2065 * Restore delay chain settings before letting the loop
2066 * in rw_mgr_mem_calibrate_vfifo to retry different
2067 * dqs/ck relationships.
2069 scc_mgr_set_dqs_bus_in_delay(read_group, start_dqs);
2070 if (IO_SHIFT_DQS_EN_WHEN_SHIFT_DQS) {
2071 scc_mgr_set_dqs_en_delay(read_group,
2074 scc_mgr_load_dqs(read_group);
2075 writel(0, &sdr_scc_mgr->update);
2077 debug_cond(DLEVEL == 1, "%s:%d vfifo_center: failed to \
2078 find edge [%u]: %d %d", __func__, __LINE__,
2079 i, left_edge[i], right_edge[i]);
2080 if (use_read_test) {
2081 set_failing_group_stage(read_group *
2082 RW_MGR_MEM_DQ_PER_READ_DQS + i,
2084 CAL_SUBSTAGE_VFIFO_CENTER);
2086 set_failing_group_stage(read_group *
2087 RW_MGR_MEM_DQ_PER_READ_DQS + i,
2088 CAL_STAGE_VFIFO_AFTER_WRITES,
2089 CAL_SUBSTAGE_VFIFO_CENTER);
2095 /* Find middle of window for each DQ bit */
2096 mid_min = left_edge[0] - right_edge[0];
2098 for (i = 1; i < RW_MGR_MEM_DQ_PER_READ_DQS; i++) {
2099 mid = left_edge[i] - right_edge[i];
2100 if (mid < mid_min) {
2107 * -mid_min/2 represents the amount that we need to move DQS.
2108 * If mid_min is odd and positive we'll need to add one to
2109 * make sure the rounding in further calculations is correct
2110 * (always bias to the right), so just add 1 for all positive values.
2115 mid_min = mid_min / 2;
2117 debug_cond(DLEVEL == 1, "%s:%d vfifo_center: mid_min=%d (index=%u)\n",
2118 __func__, __LINE__, mid_min, min_index);
2120 /* Determine the amount we can change DQS (which is -mid_min) */
2121 orig_mid_min = mid_min;
2122 new_dqs = start_dqs - mid_min;
2123 if (new_dqs > IO_DQS_IN_DELAY_MAX)
2124 new_dqs = IO_DQS_IN_DELAY_MAX;
2125 else if (new_dqs < 0)
2128 mid_min = start_dqs - new_dqs;
2129 debug_cond(DLEVEL == 1, "vfifo_center: new mid_min=%d new_dqs=%d\n",
2132 if (IO_SHIFT_DQS_EN_WHEN_SHIFT_DQS) {
2133 if (start_dqs_en - mid_min > IO_DQS_EN_DELAY_MAX)
2134 mid_min += start_dqs_en - mid_min - IO_DQS_EN_DELAY_MAX;
2135 else if (start_dqs_en - mid_min < 0)
2136 mid_min += start_dqs_en - mid_min;
2138 new_dqs = start_dqs - mid_min;
2140 debug_cond(DLEVEL == 1, "vfifo_center: start_dqs=%d start_dqs_en=%d \
2141 new_dqs=%d mid_min=%d\n", start_dqs,
2142 IO_SHIFT_DQS_EN_WHEN_SHIFT_DQS ? start_dqs_en : -1,
2145 /* Initialize data for export structures */
2146 dqs_margin = IO_IO_IN_DELAY_MAX + 1;
2147 dq_margin = IO_IO_IN_DELAY_MAX + 1;
2149 /* add delay to bring centre of all DQ windows to the same "level" */
2150 for (i = 0, p = test_bgn; i < RW_MGR_MEM_DQ_PER_READ_DQS; i++, p++) {
2151 /* Use values before divide by 2 to reduce round off error */
2152 shift_dq = (left_edge[i] - right_edge[i] -
2153 (left_edge[min_index] - right_edge[min_index]))/2 +
2154 (orig_mid_min - mid_min);
2156 debug_cond(DLEVEL == 2, "vfifo_center: before: \
2157 shift_dq[%u]=%d\n", i, shift_dq);
2159 addr = SDR_PHYGRP_SCCGRP_ADDRESS | SCC_MGR_IO_IN_DELAY_OFFSET;
2160 temp_dq_in_delay1 = readl(addr + (p << 2));
2161 temp_dq_in_delay2 = readl(addr + (i << 2));
2163 if (shift_dq + (int32_t)temp_dq_in_delay1 >
2164 (int32_t)IO_IO_IN_DELAY_MAX) {
2165 shift_dq = (int32_t)IO_IO_IN_DELAY_MAX - temp_dq_in_delay2;
2166 } else if (shift_dq + (int32_t)temp_dq_in_delay1 < 0) {
2167 shift_dq = -(int32_t)temp_dq_in_delay1;
2169 debug_cond(DLEVEL == 2, "vfifo_center: after: \
2170 shift_dq[%u]=%d\n", i, shift_dq);
2171 final_dq[i] = temp_dq_in_delay1 + shift_dq;
2172 scc_mgr_set_dq_in_delay(p, final_dq[i]);
2175 debug_cond(DLEVEL == 2, "vfifo_center: margin[%u]=[%d,%d]\n", i,
2176 left_edge[i] - shift_dq + (-mid_min),
2177 right_edge[i] + shift_dq - (-mid_min));
2178 /* To determine values for export structures */
2179 if (left_edge[i] - shift_dq + (-mid_min) < dq_margin)
2180 dq_margin = left_edge[i] - shift_dq + (-mid_min);
2182 if (right_edge[i] + shift_dq - (-mid_min) < dqs_margin)
2183 dqs_margin = right_edge[i] + shift_dq - (-mid_min);
2186 final_dqs = new_dqs;
2187 if (IO_SHIFT_DQS_EN_WHEN_SHIFT_DQS)
2188 final_dqs_en = start_dqs_en - mid_min;
2191 if (IO_SHIFT_DQS_EN_WHEN_SHIFT_DQS) {
2192 scc_mgr_set_dqs_en_delay(read_group, final_dqs_en);
2193 scc_mgr_load_dqs(read_group);
2197 scc_mgr_set_dqs_bus_in_delay(read_group, final_dqs);
2198 scc_mgr_load_dqs(read_group);
2199 debug_cond(DLEVEL == 2, "%s:%d vfifo_center: dq_margin=%d \
2200 dqs_margin=%d", __func__, __LINE__,
2201 dq_margin, dqs_margin);
2204 * Do not remove this line as it makes sure all of our decisions
2205 * have been applied. Apply the update bit.
2207 writel(0, &sdr_scc_mgr->update);
2209 return (dq_margin >= 0) && (dqs_margin >= 0);
2213 * calibrate the read valid prediction FIFO.
2215 * - read valid prediction will consist of finding a good DQS enable phase,
2216 * DQS enable delay, DQS input phase, and DQS input delay.
2217 * - we also do a per-bit deskew on the DQ lines.
2219 static uint32_t rw_mgr_mem_calibrate_vfifo(uint32_t read_group,
2222 uint32_t p, d, rank_bgn, sr;
2223 uint32_t dtaps_per_ptap;
2226 uint32_t grp_calibrated;
2227 uint32_t write_group, write_test_bgn;
2228 uint32_t failed_substage;
2230 debug("%s:%d: %u %u\n", __func__, __LINE__, read_group, test_bgn);
2232 /* update info for sims */
2233 reg_file_set_stage(CAL_STAGE_VFIFO);
2235 write_group = read_group;
2236 write_test_bgn = test_bgn;
2238 /* USER Determine number of delay taps for each phase tap */
2241 while (tmp_delay < IO_DELAY_PER_OPA_TAP) {
2243 tmp_delay += IO_DELAY_PER_DQS_EN_DCHAIN_TAP;
2248 /* update info for sims */
2249 reg_file_set_group(read_group);
2253 reg_file_set_sub_stage(CAL_SUBSTAGE_GUARANTEED_READ);
2254 failed_substage = CAL_SUBSTAGE_GUARANTEED_READ;
2256 for (d = 0; d <= dtaps_per_ptap && grp_calibrated == 0; d += 2) {
2258 * In RLDRAMX we may be messing the delay of pins in
2259 * the same write group but outside of the current read
2260 * the group, but that's ok because we haven't
2261 * calibrated output side yet.
2264 scc_mgr_apply_group_all_out_delay_add_all_ranks
2265 (write_group, write_test_bgn, d);
2268 for (p = 0; p <= IO_DQDQS_OUT_PHASE_MAX && grp_calibrated == 0;
2270 /* set a particular dqdqs phase */
2271 scc_mgr_set_dqdqs_output_phase_all_ranks(read_group, p);
2273 debug_cond(DLEVEL == 1, "%s:%d calibrate_vfifo: g=%u \
2274 p=%u d=%u\n", __func__, __LINE__,
2278 * Load up the patterns used by read calibration
2279 * using current DQDQS phase.
2281 rw_mgr_mem_calibrate_read_load_patterns(0, 1);
2282 if (!(gbl->phy_debug_mode_flags &
2283 PHY_DEBUG_DISABLE_GUARANTEED_READ)) {
2284 if (!rw_mgr_mem_calibrate_read_test_patterns_all_ranks
2285 (read_group, 1, &bit_chk)) {
2286 debug_cond(DLEVEL == 1, "%s:%d Guaranteed read test failed:",
2287 __func__, __LINE__);
2288 debug_cond(DLEVEL == 1, " g=%u p=%u d=%u\n",
2296 if (rw_mgr_mem_calibrate_vfifo_find_dqs_en_phase_sweep_dq_in_delay
2297 (write_group, read_group, test_bgn)) {
2299 * USER Read per-bit deskew can be done on a
2300 * per shadow register basis.
2302 for (rank_bgn = 0, sr = 0;
2303 rank_bgn < RW_MGR_MEM_NUMBER_OF_RANKS;
2304 rank_bgn += NUM_RANKS_PER_SHADOW_REG,
2307 * Determine if this set of ranks
2308 * should be skipped entirely.
2310 if (!param->skip_shadow_regs[sr]) {
2312 * If doing read after write
2313 * calibration, do not update
2314 * FOM, now - do it then.
2316 if (!rw_mgr_mem_calibrate_vfifo_center
2317 (rank_bgn, write_group,
2318 read_group, test_bgn, 1, 0)) {
2321 CAL_SUBSTAGE_VFIFO_CENTER;
2327 failed_substage = CAL_SUBSTAGE_DQS_EN_PHASE;
2332 if (grp_calibrated == 0) {
2333 set_failing_group_stage(write_group, CAL_STAGE_VFIFO,
2339 * Reset the delay chains back to zero if they have moved > 1
2340 * (check for > 1 because loop will increase d even when pass in
2344 scc_mgr_zero_group(write_group, write_test_bgn, 1);
2349 /* VFIFO Calibration -- Read Deskew Calibration after write deskew */
2350 static uint32_t rw_mgr_mem_calibrate_vfifo_end(uint32_t read_group,
2353 uint32_t rank_bgn, sr;
2354 uint32_t grp_calibrated;
2355 uint32_t write_group;
2357 debug("%s:%d %u %u", __func__, __LINE__, read_group, test_bgn);
2359 /* update info for sims */
2361 reg_file_set_stage(CAL_STAGE_VFIFO_AFTER_WRITES);
2362 reg_file_set_sub_stage(CAL_SUBSTAGE_VFIFO_CENTER);
2364 write_group = read_group;
2366 /* update info for sims */
2367 reg_file_set_group(read_group);
2370 /* Read per-bit deskew can be done on a per shadow register basis */
2371 for (rank_bgn = 0, sr = 0; rank_bgn < RW_MGR_MEM_NUMBER_OF_RANKS;
2372 rank_bgn += NUM_RANKS_PER_SHADOW_REG, ++sr) {
2373 /* Determine if this set of ranks should be skipped entirely */
2374 if (!param->skip_shadow_regs[sr]) {
2375 /* This is the last calibration round, update FOM here */
2376 if (!rw_mgr_mem_calibrate_vfifo_center(rank_bgn,
2387 if (grp_calibrated == 0) {
2388 set_failing_group_stage(write_group,
2389 CAL_STAGE_VFIFO_AFTER_WRITES,
2390 CAL_SUBSTAGE_VFIFO_CENTER);
2397 /* Calibrate LFIFO to find smallest read latency */
2398 static uint32_t rw_mgr_mem_calibrate_lfifo(void)
2403 debug("%s:%d\n", __func__, __LINE__);
2405 /* update info for sims */
2406 reg_file_set_stage(CAL_STAGE_LFIFO);
2407 reg_file_set_sub_stage(CAL_SUBSTAGE_READ_LATENCY);
2409 /* Load up the patterns used by read calibration for all ranks */
2410 rw_mgr_mem_calibrate_read_load_patterns(0, 1);
2414 writel(gbl->curr_read_lat, &phy_mgr_cfg->phy_rlat);
2415 debug_cond(DLEVEL == 2, "%s:%d lfifo: read_lat=%u",
2416 __func__, __LINE__, gbl->curr_read_lat);
2418 if (!rw_mgr_mem_calibrate_read_test_all_ranks(0,
2426 /* reduce read latency and see if things are working */
2428 gbl->curr_read_lat--;
2429 } while (gbl->curr_read_lat > 0);
2431 /* reset the fifos to get pointers to known state */
2433 writel(0, &phy_mgr_cmd->fifo_reset);
2436 /* add a fudge factor to the read latency that was determined */
2437 gbl->curr_read_lat += 2;
2438 writel(gbl->curr_read_lat, &phy_mgr_cfg->phy_rlat);
2439 debug_cond(DLEVEL == 2, "%s:%d lfifo: success: using \
2440 read_lat=%u\n", __func__, __LINE__,
2441 gbl->curr_read_lat);
2444 set_failing_group_stage(0xff, CAL_STAGE_LFIFO,
2445 CAL_SUBSTAGE_READ_LATENCY);
2447 debug_cond(DLEVEL == 2, "%s:%d lfifo: failed at initial \
2448 read_lat=%u\n", __func__, __LINE__,
2449 gbl->curr_read_lat);
2455 * issue write test command.
2456 * two variants are provided. one that just tests a write pattern and
2457 * another that tests datamask functionality.
2459 static void rw_mgr_mem_calibrate_write_test_issue(uint32_t group,
2462 uint32_t mcc_instruction;
2463 uint32_t quick_write_mode = (((STATIC_CALIB_STEPS) & CALIB_SKIP_WRITES) &&
2464 ENABLE_SUPER_QUICK_CALIBRATION);
2465 uint32_t rw_wl_nop_cycles;
2469 * Set counter and jump addresses for the right
2470 * number of NOP cycles.
2471 * The number of supported NOP cycles can range from -1 to infinity
2472 * Three different cases are handled:
2474 * 1. For a number of NOP cycles greater than 0, the RW Mgr looping
2475 * mechanism will be used to insert the right number of NOPs
2477 * 2. For a number of NOP cycles equals to 0, the micro-instruction
2478 * issuing the write command will jump straight to the
2479 * micro-instruction that turns on DQS (for DDRx), or outputs write
2480 * data (for RLD), skipping
2481 * the NOP micro-instruction all together
2483 * 3. A number of NOP cycles equal to -1 indicates that DQS must be
2484 * turned on in the same micro-instruction that issues the write
2485 * command. Then we need
2486 * to directly jump to the micro-instruction that sends out the data
2488 * NOTE: Implementing this mechanism uses 2 RW Mgr jump-counters
2489 * (2 and 3). One jump-counter (0) is used to perform multiple
2490 * write-read operations.
2491 * one counter left to issue this command in "multiple-group" mode
2494 rw_wl_nop_cycles = gbl->rw_wl_nop_cycles;
2496 if (rw_wl_nop_cycles == -1) {
2498 * CNTR 2 - We want to execute the special write operation that
2499 * turns on DQS right away and then skip directly to the
2500 * instruction that sends out the data. We set the counter to a
2501 * large number so that the jump is always taken.
2503 writel(0xFF, &sdr_rw_load_mgr_regs->load_cntr2);
2505 /* CNTR 3 - Not used */
2507 mcc_instruction = RW_MGR_LFSR_WR_RD_DM_BANK_0_WL_1;
2508 writel(RW_MGR_LFSR_WR_RD_DM_BANK_0_DATA,
2509 &sdr_rw_load_jump_mgr_regs->load_jump_add2);
2510 writel(RW_MGR_LFSR_WR_RD_DM_BANK_0_NOP,
2511 &sdr_rw_load_jump_mgr_regs->load_jump_add3);
2513 mcc_instruction = RW_MGR_LFSR_WR_RD_BANK_0_WL_1;
2514 writel(RW_MGR_LFSR_WR_RD_BANK_0_DATA,
2515 &sdr_rw_load_jump_mgr_regs->load_jump_add2);
2516 writel(RW_MGR_LFSR_WR_RD_BANK_0_NOP,
2517 &sdr_rw_load_jump_mgr_regs->load_jump_add3);
2519 } else if (rw_wl_nop_cycles == 0) {
2521 * CNTR 2 - We want to skip the NOP operation and go straight
2522 * to the DQS enable instruction. We set the counter to a large
2523 * number so that the jump is always taken.
2525 writel(0xFF, &sdr_rw_load_mgr_regs->load_cntr2);
2527 /* CNTR 3 - Not used */
2529 mcc_instruction = RW_MGR_LFSR_WR_RD_DM_BANK_0;
2530 writel(RW_MGR_LFSR_WR_RD_DM_BANK_0_DQS,
2531 &sdr_rw_load_jump_mgr_regs->load_jump_add2);
2533 mcc_instruction = RW_MGR_LFSR_WR_RD_BANK_0;
2534 writel(RW_MGR_LFSR_WR_RD_BANK_0_DQS,
2535 &sdr_rw_load_jump_mgr_regs->load_jump_add2);
2539 * CNTR 2 - In this case we want to execute the next instruction
2540 * and NOT take the jump. So we set the counter to 0. The jump
2541 * address doesn't count.
2543 writel(0x0, &sdr_rw_load_mgr_regs->load_cntr2);
2544 writel(0x0, &sdr_rw_load_jump_mgr_regs->load_jump_add2);
2547 * CNTR 3 - Set the nop counter to the number of cycles we
2548 * need to loop for, minus 1.
2550 writel(rw_wl_nop_cycles - 1, &sdr_rw_load_mgr_regs->load_cntr3);
2552 mcc_instruction = RW_MGR_LFSR_WR_RD_DM_BANK_0;
2553 writel(RW_MGR_LFSR_WR_RD_DM_BANK_0_NOP,
2554 &sdr_rw_load_jump_mgr_regs->load_jump_add3);
2556 mcc_instruction = RW_MGR_LFSR_WR_RD_BANK_0;
2557 writel(RW_MGR_LFSR_WR_RD_BANK_0_NOP,
2558 &sdr_rw_load_jump_mgr_regs->load_jump_add3);
2562 writel(0, SDR_PHYGRP_RWMGRGRP_ADDRESS |
2563 RW_MGR_RESET_READ_DATAPATH_OFFSET);
2565 if (quick_write_mode)
2566 writel(0x08, &sdr_rw_load_mgr_regs->load_cntr0);
2568 writel(0x40, &sdr_rw_load_mgr_regs->load_cntr0);
2570 writel(mcc_instruction, &sdr_rw_load_jump_mgr_regs->load_jump_add0);
2573 * CNTR 1 - This is used to ensure enough time elapses
2574 * for read data to come back.
2576 writel(0x30, &sdr_rw_load_mgr_regs->load_cntr1);
2579 writel(RW_MGR_LFSR_WR_RD_DM_BANK_0_WAIT,
2580 &sdr_rw_load_jump_mgr_regs->load_jump_add1);
2582 writel(RW_MGR_LFSR_WR_RD_BANK_0_WAIT,
2583 &sdr_rw_load_jump_mgr_regs->load_jump_add1);
2586 addr = SDR_PHYGRP_RWMGRGRP_ADDRESS | RW_MGR_RUN_SINGLE_GROUP_OFFSET;
2587 writel(mcc_instruction, addr + (group << 2));
2590 /* Test writes, can check for a single bit pass or multiple bit pass */
2591 static uint32_t rw_mgr_mem_calibrate_write_test(uint32_t rank_bgn,
2592 uint32_t write_group, uint32_t use_dm, uint32_t all_correct,
2593 uint32_t *bit_chk, uint32_t all_ranks)
2596 uint32_t correct_mask_vg;
2597 uint32_t tmp_bit_chk;
2599 uint32_t rank_end = all_ranks ? RW_MGR_MEM_NUMBER_OF_RANKS :
2600 (rank_bgn + NUM_RANKS_PER_SHADOW_REG);
2601 uint32_t addr_rw_mgr;
2602 uint32_t base_rw_mgr;
2604 *bit_chk = param->write_correct_mask;
2605 correct_mask_vg = param->write_correct_mask_vg;
2607 for (r = rank_bgn; r < rank_end; r++) {
2608 if (param->skip_ranks[r]) {
2609 /* request to skip the rank */
2614 set_rank_and_odt_mask(r, RW_MGR_ODT_MODE_READ_WRITE);
2617 addr_rw_mgr = SDR_PHYGRP_RWMGRGRP_ADDRESS;
2618 for (vg = RW_MGR_MEM_VIRTUAL_GROUPS_PER_WRITE_DQS-1; ; vg--) {
2619 /* reset the fifos to get pointers to known state */
2620 writel(0, &phy_mgr_cmd->fifo_reset);
2622 tmp_bit_chk = tmp_bit_chk <<
2623 (RW_MGR_MEM_DQ_PER_WRITE_DQS /
2624 RW_MGR_MEM_VIRTUAL_GROUPS_PER_WRITE_DQS);
2625 rw_mgr_mem_calibrate_write_test_issue(write_group *
2626 RW_MGR_MEM_VIRTUAL_GROUPS_PER_WRITE_DQS+vg,
2629 base_rw_mgr = readl(addr_rw_mgr);
2630 tmp_bit_chk = tmp_bit_chk | (correct_mask_vg & ~(base_rw_mgr));
2634 *bit_chk &= tmp_bit_chk;
2638 set_rank_and_odt_mask(0, RW_MGR_ODT_MODE_OFF);
2639 debug_cond(DLEVEL == 2, "write_test(%u,%u,ALL) : %u == \
2640 %u => %lu", write_group, use_dm,
2641 *bit_chk, param->write_correct_mask,
2642 (long unsigned int)(*bit_chk ==
2643 param->write_correct_mask));
2644 return *bit_chk == param->write_correct_mask;
2646 set_rank_and_odt_mask(0, RW_MGR_ODT_MODE_OFF);
2647 debug_cond(DLEVEL == 2, "write_test(%u,%u,ONE) : %u != ",
2648 write_group, use_dm, *bit_chk);
2649 debug_cond(DLEVEL == 2, "%lu" " => %lu", (long unsigned int)0,
2650 (long unsigned int)(*bit_chk != 0));
2651 return *bit_chk != 0x00;
2656 * center all windows. do per-bit-deskew to possibly increase size of
2659 static uint32_t rw_mgr_mem_calibrate_writes_center(uint32_t rank_bgn,
2660 uint32_t write_group, uint32_t test_bgn)
2662 uint32_t i, p, min_index;
2665 * Store these as signed since there are comparisons with
2669 uint32_t sticky_bit_chk;
2670 int32_t left_edge[RW_MGR_MEM_DQ_PER_WRITE_DQS];
2671 int32_t right_edge[RW_MGR_MEM_DQ_PER_WRITE_DQS];
2673 int32_t mid_min, orig_mid_min;
2674 int32_t new_dqs, start_dqs, shift_dq;
2675 int32_t dq_margin, dqs_margin, dm_margin;
2677 uint32_t temp_dq_out1_delay;
2680 debug("%s:%d %u %u", __func__, __LINE__, write_group, test_bgn);
2684 addr = SDR_PHYGRP_SCCGRP_ADDRESS | SCC_MGR_IO_OUT1_DELAY_OFFSET;
2685 start_dqs = readl(addr +
2686 (RW_MGR_MEM_DQ_PER_WRITE_DQS << 2));
2688 /* per-bit deskew */
2691 * set the left and right edge of each bit to an illegal value
2692 * use (IO_IO_OUT1_DELAY_MAX + 1) as an illegal value.
2695 for (i = 0; i < RW_MGR_MEM_DQ_PER_WRITE_DQS; i++) {
2696 left_edge[i] = IO_IO_OUT1_DELAY_MAX + 1;
2697 right_edge[i] = IO_IO_OUT1_DELAY_MAX + 1;
2700 /* Search for the left edge of the window for each bit */
2701 for (d = 0; d <= IO_IO_OUT1_DELAY_MAX; d++) {
2702 scc_mgr_apply_group_dq_out1_delay(write_group, test_bgn, d);
2704 writel(0, &sdr_scc_mgr->update);
2707 * Stop searching when the read test doesn't pass AND when
2708 * we've seen a passing read on every bit.
2710 stop = !rw_mgr_mem_calibrate_write_test(rank_bgn, write_group,
2711 0, PASS_ONE_BIT, &bit_chk, 0);
2712 sticky_bit_chk = sticky_bit_chk | bit_chk;
2713 stop = stop && (sticky_bit_chk == param->write_correct_mask);
2714 debug_cond(DLEVEL == 2, "write_center(left): dtap=%d => %u \
2715 == %u && %u [bit_chk= %u ]\n",
2716 d, sticky_bit_chk, param->write_correct_mask,
2722 for (i = 0; i < RW_MGR_MEM_DQ_PER_WRITE_DQS; i++) {
2725 * Remember a passing test as the
2731 * If a left edge has not been seen
2732 * yet, then a future passing test will
2733 * mark this edge as the right edge.
2736 IO_IO_OUT1_DELAY_MAX + 1) {
2737 right_edge[i] = -(d + 1);
2740 debug_cond(DLEVEL == 2, "write_center[l,d=%d):", d);
2741 debug_cond(DLEVEL == 2, "bit_chk_test=%d left_edge[%u]: %d",
2742 (int)(bit_chk & 1), i, left_edge[i]);
2743 debug_cond(DLEVEL == 2, "right_edge[%u]: %d\n", i,
2745 bit_chk = bit_chk >> 1;
2750 /* Reset DQ delay chains to 0 */
2751 scc_mgr_apply_group_dq_out1_delay(write_group, test_bgn, 0);
2753 for (i = RW_MGR_MEM_DQ_PER_WRITE_DQS - 1;; i--) {
2754 debug_cond(DLEVEL == 2, "%s:%d write_center: left_edge[%u]: \
2755 %d right_edge[%u]: %d\n", __func__, __LINE__,
2756 i, left_edge[i], i, right_edge[i]);
2759 * Check for cases where we haven't found the left edge,
2760 * which makes our assignment of the the right edge invalid.
2761 * Reset it to the illegal value.
2763 if ((left_edge[i] == IO_IO_OUT1_DELAY_MAX + 1) &&
2764 (right_edge[i] != IO_IO_OUT1_DELAY_MAX + 1)) {
2765 right_edge[i] = IO_IO_OUT1_DELAY_MAX + 1;
2766 debug_cond(DLEVEL == 2, "%s:%d write_center: reset \
2767 right_edge[%u]: %d\n", __func__, __LINE__,
2772 * Reset sticky bit (except for bits where we have
2773 * seen the left edge).
2775 sticky_bit_chk = sticky_bit_chk << 1;
2776 if ((left_edge[i] != IO_IO_OUT1_DELAY_MAX + 1))
2777 sticky_bit_chk = sticky_bit_chk | 1;
2783 /* Search for the right edge of the window for each bit */
2784 for (d = 0; d <= IO_IO_OUT1_DELAY_MAX - start_dqs; d++) {
2785 scc_mgr_apply_group_dqs_io_and_oct_out1(write_group,
2788 writel(0, &sdr_scc_mgr->update);
2791 * Stop searching when the read test doesn't pass AND when
2792 * we've seen a passing read on every bit.
2794 stop = !rw_mgr_mem_calibrate_write_test(rank_bgn, write_group,
2795 0, PASS_ONE_BIT, &bit_chk, 0);
2797 sticky_bit_chk = sticky_bit_chk | bit_chk;
2798 stop = stop && (sticky_bit_chk == param->write_correct_mask);
2800 debug_cond(DLEVEL == 2, "write_center (right): dtap=%u => %u == \
2801 %u && %u\n", d, sticky_bit_chk,
2802 param->write_correct_mask, stop);
2806 for (i = 0; i < RW_MGR_MEM_DQ_PER_WRITE_DQS;
2808 /* d = 0 failed, but it passed when
2809 testing the left edge, so it must be
2810 marginal, set it to -1 */
2811 if (right_edge[i] ==
2812 IO_IO_OUT1_DELAY_MAX + 1 &&
2814 IO_IO_OUT1_DELAY_MAX + 1) {
2821 for (i = 0; i < RW_MGR_MEM_DQ_PER_WRITE_DQS; i++) {
2824 * Remember a passing test as
2831 * If a right edge has not
2832 * been seen yet, then a future
2833 * passing test will mark this
2834 * edge as the left edge.
2836 if (right_edge[i] ==
2837 IO_IO_OUT1_DELAY_MAX + 1)
2838 left_edge[i] = -(d + 1);
2841 * d = 0 failed, but it passed
2842 * when testing the left edge,
2843 * so it must be marginal, set
2846 if (right_edge[i] ==
2847 IO_IO_OUT1_DELAY_MAX + 1 &&
2849 IO_IO_OUT1_DELAY_MAX + 1)
2852 * If a right edge has not been
2853 * seen yet, then a future
2854 * passing test will mark this
2855 * edge as the left edge.
2857 else if (right_edge[i] ==
2858 IO_IO_OUT1_DELAY_MAX +
2860 left_edge[i] = -(d + 1);
2863 debug_cond(DLEVEL == 2, "write_center[r,d=%d):", d);
2864 debug_cond(DLEVEL == 2, "bit_chk_test=%d left_edge[%u]: %d",
2865 (int)(bit_chk & 1), i, left_edge[i]);
2866 debug_cond(DLEVEL == 2, "right_edge[%u]: %d\n", i,
2868 bit_chk = bit_chk >> 1;
2873 /* Check that all bits have a window */
2874 for (i = 0; i < RW_MGR_MEM_DQ_PER_WRITE_DQS; i++) {
2875 debug_cond(DLEVEL == 2, "%s:%d write_center: left_edge[%u]: \
2876 %d right_edge[%u]: %d", __func__, __LINE__,
2877 i, left_edge[i], i, right_edge[i]);
2878 if ((left_edge[i] == IO_IO_OUT1_DELAY_MAX + 1) ||
2879 (right_edge[i] == IO_IO_OUT1_DELAY_MAX + 1)) {
2880 set_failing_group_stage(test_bgn + i,
2882 CAL_SUBSTAGE_WRITES_CENTER);
2887 /* Find middle of window for each DQ bit */
2888 mid_min = left_edge[0] - right_edge[0];
2890 for (i = 1; i < RW_MGR_MEM_DQ_PER_WRITE_DQS; i++) {
2891 mid = left_edge[i] - right_edge[i];
2892 if (mid < mid_min) {
2899 * -mid_min/2 represents the amount that we need to move DQS.
2900 * If mid_min is odd and positive we'll need to add one to
2901 * make sure the rounding in further calculations is correct
2902 * (always bias to the right), so just add 1 for all positive values.
2906 mid_min = mid_min / 2;
2907 debug_cond(DLEVEL == 1, "%s:%d write_center: mid_min=%d\n", __func__,
2910 /* Determine the amount we can change DQS (which is -mid_min) */
2911 orig_mid_min = mid_min;
2912 new_dqs = start_dqs;
2914 debug_cond(DLEVEL == 1, "%s:%d write_center: start_dqs=%d new_dqs=%d \
2915 mid_min=%d\n", __func__, __LINE__, start_dqs, new_dqs, mid_min);
2916 /* Initialize data for export structures */
2917 dqs_margin = IO_IO_OUT1_DELAY_MAX + 1;
2918 dq_margin = IO_IO_OUT1_DELAY_MAX + 1;
2920 /* add delay to bring centre of all DQ windows to the same "level" */
2921 for (i = 0, p = test_bgn; i < RW_MGR_MEM_DQ_PER_WRITE_DQS; i++, p++) {
2922 /* Use values before divide by 2 to reduce round off error */
2923 shift_dq = (left_edge[i] - right_edge[i] -
2924 (left_edge[min_index] - right_edge[min_index]))/2 +
2925 (orig_mid_min - mid_min);
2927 debug_cond(DLEVEL == 2, "%s:%d write_center: before: shift_dq \
2928 [%u]=%d\n", __func__, __LINE__, i, shift_dq);
2930 addr = SDR_PHYGRP_SCCGRP_ADDRESS | SCC_MGR_IO_OUT1_DELAY_OFFSET;
2931 temp_dq_out1_delay = readl(addr + (i << 2));
2932 if (shift_dq + (int32_t)temp_dq_out1_delay >
2933 (int32_t)IO_IO_OUT1_DELAY_MAX) {
2934 shift_dq = (int32_t)IO_IO_OUT1_DELAY_MAX - temp_dq_out1_delay;
2935 } else if (shift_dq + (int32_t)temp_dq_out1_delay < 0) {
2936 shift_dq = -(int32_t)temp_dq_out1_delay;
2938 debug_cond(DLEVEL == 2, "write_center: after: shift_dq[%u]=%d\n",
2940 scc_mgr_set_dq_out1_delay(i, temp_dq_out1_delay + shift_dq);
2943 debug_cond(DLEVEL == 2, "write_center: margin[%u]=[%d,%d]\n", i,
2944 left_edge[i] - shift_dq + (-mid_min),
2945 right_edge[i] + shift_dq - (-mid_min));
2946 /* To determine values for export structures */
2947 if (left_edge[i] - shift_dq + (-mid_min) < dq_margin)
2948 dq_margin = left_edge[i] - shift_dq + (-mid_min);
2950 if (right_edge[i] + shift_dq - (-mid_min) < dqs_margin)
2951 dqs_margin = right_edge[i] + shift_dq - (-mid_min);
2955 scc_mgr_apply_group_dqs_io_and_oct_out1(write_group, new_dqs);
2956 writel(0, &sdr_scc_mgr->update);
2959 debug_cond(DLEVEL == 2, "%s:%d write_center: DM\n", __func__, __LINE__);
2962 * set the left and right edge of each bit to an illegal value,
2963 * use (IO_IO_OUT1_DELAY_MAX + 1) as an illegal value,
2965 left_edge[0] = IO_IO_OUT1_DELAY_MAX + 1;
2966 right_edge[0] = IO_IO_OUT1_DELAY_MAX + 1;
2967 int32_t bgn_curr = IO_IO_OUT1_DELAY_MAX + 1;
2968 int32_t end_curr = IO_IO_OUT1_DELAY_MAX + 1;
2969 int32_t bgn_best = IO_IO_OUT1_DELAY_MAX + 1;
2970 int32_t end_best = IO_IO_OUT1_DELAY_MAX + 1;
2971 int32_t win_best = 0;
2973 /* Search for the/part of the window with DM shift */
2974 for (d = IO_IO_OUT1_DELAY_MAX; d >= 0; d -= DELTA_D) {
2975 scc_mgr_apply_group_dm_out1_delay(write_group, d);
2976 writel(0, &sdr_scc_mgr->update);
2978 if (rw_mgr_mem_calibrate_write_test(rank_bgn, write_group, 1,
2979 PASS_ALL_BITS, &bit_chk,
2981 /* USE Set current end of the window */
2984 * If a starting edge of our window has not been seen
2985 * this is our current start of the DM window.
2987 if (bgn_curr == IO_IO_OUT1_DELAY_MAX + 1)
2991 * If current window is bigger than best seen.
2992 * Set best seen to be current window.
2994 if ((end_curr-bgn_curr+1) > win_best) {
2995 win_best = end_curr-bgn_curr+1;
2996 bgn_best = bgn_curr;
2997 end_best = end_curr;
3000 /* We just saw a failing test. Reset temp edge */
3001 bgn_curr = IO_IO_OUT1_DELAY_MAX + 1;
3002 end_curr = IO_IO_OUT1_DELAY_MAX + 1;
3007 /* Reset DM delay chains to 0 */
3008 scc_mgr_apply_group_dm_out1_delay(write_group, 0);
3011 * Check to see if the current window nudges up aganist 0 delay.
3012 * If so we need to continue the search by shifting DQS otherwise DQS
3013 * search begins as a new search. */
3014 if (end_curr != 0) {
3015 bgn_curr = IO_IO_OUT1_DELAY_MAX + 1;
3016 end_curr = IO_IO_OUT1_DELAY_MAX + 1;
3019 /* Search for the/part of the window with DQS shifts */
3020 for (d = 0; d <= IO_IO_OUT1_DELAY_MAX - new_dqs; d += DELTA_D) {
3022 * Note: This only shifts DQS, so are we limiting ourselve to
3023 * width of DQ unnecessarily.
3025 scc_mgr_apply_group_dqs_io_and_oct_out1(write_group,
3028 writel(0, &sdr_scc_mgr->update);
3029 if (rw_mgr_mem_calibrate_write_test(rank_bgn, write_group, 1,
3030 PASS_ALL_BITS, &bit_chk,
3032 /* USE Set current end of the window */
3035 * If a beginning edge of our window has not been seen
3036 * this is our current begin of the DM window.
3038 if (bgn_curr == IO_IO_OUT1_DELAY_MAX + 1)
3042 * If current window is bigger than best seen. Set best
3043 * seen to be current window.
3045 if ((end_curr-bgn_curr+1) > win_best) {
3046 win_best = end_curr-bgn_curr+1;
3047 bgn_best = bgn_curr;
3048 end_best = end_curr;
3051 /* We just saw a failing test. Reset temp edge */
3052 bgn_curr = IO_IO_OUT1_DELAY_MAX + 1;
3053 end_curr = IO_IO_OUT1_DELAY_MAX + 1;
3055 /* Early exit optimization: if ther remaining delay
3056 chain space is less than already seen largest window
3059 (IO_IO_OUT1_DELAY_MAX - new_dqs - d)) {
3065 /* assign left and right edge for cal and reporting; */
3066 left_edge[0] = -1*bgn_best;
3067 right_edge[0] = end_best;
3069 debug_cond(DLEVEL == 2, "%s:%d dm_calib: left=%d right=%d\n", __func__,
3070 __LINE__, left_edge[0], right_edge[0]);
3072 /* Move DQS (back to orig) */
3073 scc_mgr_apply_group_dqs_io_and_oct_out1(write_group, new_dqs);
3077 /* Find middle of window for the DM bit */
3078 mid = (left_edge[0] - right_edge[0]) / 2;
3080 /* only move right, since we are not moving DQS/DQ */
3084 /* dm_marign should fail if we never find a window */
3088 dm_margin = left_edge[0] - mid;
3090 scc_mgr_apply_group_dm_out1_delay(write_group, mid);
3091 writel(0, &sdr_scc_mgr->update);
3093 debug_cond(DLEVEL == 2, "%s:%d dm_calib: left=%d right=%d mid=%d \
3094 dm_margin=%d\n", __func__, __LINE__, left_edge[0],
3095 right_edge[0], mid, dm_margin);
3097 gbl->fom_out += dq_margin + dqs_margin;
3099 debug_cond(DLEVEL == 2, "%s:%d write_center: dq_margin=%d \
3100 dqs_margin=%d dm_margin=%d\n", __func__, __LINE__,
3101 dq_margin, dqs_margin, dm_margin);
3104 * Do not remove this line as it makes sure all of our
3105 * decisions have been applied.
3107 writel(0, &sdr_scc_mgr->update);
3108 return (dq_margin >= 0) && (dqs_margin >= 0) && (dm_margin >= 0);
3111 /* calibrate the write operations */
3112 static uint32_t rw_mgr_mem_calibrate_writes(uint32_t rank_bgn, uint32_t g,
3115 /* update info for sims */
3116 debug("%s:%d %u %u\n", __func__, __LINE__, g, test_bgn);
3118 reg_file_set_stage(CAL_STAGE_WRITES);
3119 reg_file_set_sub_stage(CAL_SUBSTAGE_WRITES_CENTER);
3121 reg_file_set_group(g);
3123 if (!rw_mgr_mem_calibrate_writes_center(rank_bgn, g, test_bgn)) {
3124 set_failing_group_stage(g, CAL_STAGE_WRITES,
3125 CAL_SUBSTAGE_WRITES_CENTER);
3132 /* precharge all banks and activate row 0 in bank "000..." and bank "111..." */
3133 static void mem_precharge_and_activate(void)
3137 for (r = 0; r < RW_MGR_MEM_NUMBER_OF_RANKS; r++) {
3138 if (param->skip_ranks[r]) {
3139 /* request to skip the rank */
3144 set_rank_and_odt_mask(r, RW_MGR_ODT_MODE_OFF);
3146 /* precharge all banks ... */
3147 writel(RW_MGR_PRECHARGE_ALL, SDR_PHYGRP_RWMGRGRP_ADDRESS |
3148 RW_MGR_RUN_SINGLE_GROUP_OFFSET);
3150 writel(0x0F, &sdr_rw_load_mgr_regs->load_cntr0);
3151 writel(RW_MGR_ACTIVATE_0_AND_1_WAIT1,
3152 &sdr_rw_load_jump_mgr_regs->load_jump_add0);
3154 writel(0x0F, &sdr_rw_load_mgr_regs->load_cntr1);
3155 writel(RW_MGR_ACTIVATE_0_AND_1_WAIT2,
3156 &sdr_rw_load_jump_mgr_regs->load_jump_add1);
3159 writel(RW_MGR_ACTIVATE_0_AND_1, SDR_PHYGRP_RWMGRGRP_ADDRESS |
3160 RW_MGR_RUN_SINGLE_GROUP_OFFSET);
3164 /* Configure various memory related parameters. */
3165 static void mem_config(void)
3167 uint32_t rlat, wlat;
3168 uint32_t rw_wl_nop_cycles;
3169 uint32_t max_latency;
3171 debug("%s:%d\n", __func__, __LINE__);
3172 /* read in write and read latency */
3173 wlat = readl(&data_mgr->t_wl_add);
3174 wlat += readl(&data_mgr->mem_t_add);
3176 /* WL for hard phy does not include additive latency */
3179 * add addtional write latency to offset the address/command extra
3180 * clock cycle. We change the AC mux setting causing AC to be delayed
3181 * by one mem clock cycle. Only do this for DDR3
3185 rlat = readl(&data_mgr->t_rl_add);
3187 rw_wl_nop_cycles = wlat - 2;
3188 gbl->rw_wl_nop_cycles = rw_wl_nop_cycles;
3191 * For AV/CV, lfifo is hardened and always runs at full rate so
3192 * max latency in AFI clocks, used here, is correspondingly smaller.
3194 max_latency = (1<<MAX_LATENCY_COUNT_WIDTH)/1 - 1;
3195 /* configure for a burst length of 8 */
3198 /* Adjust Write Latency for Hard PHY */
3201 /* set a pretty high read latency initially */
3202 gbl->curr_read_lat = rlat + 16;
3204 if (gbl->curr_read_lat > max_latency)
3205 gbl->curr_read_lat = max_latency;
3207 writel(gbl->curr_read_lat, &phy_mgr_cfg->phy_rlat);
3209 /* advertise write latency */
3210 gbl->curr_write_lat = wlat;
3211 writel(wlat - 2, &phy_mgr_cfg->afi_wlat);
3213 /* initialize bit slips */
3214 mem_precharge_and_activate();
3217 /* Set VFIFO and LFIFO to instant-on settings in skip calibration mode */
3218 static void mem_skip_calibrate(void)
3220 uint32_t vfifo_offset;
3223 debug("%s:%d\n", __func__, __LINE__);
3224 /* Need to update every shadow register set used by the interface */
3225 for (r = 0; r < RW_MGR_MEM_NUMBER_OF_RANKS;
3226 r += NUM_RANKS_PER_SHADOW_REG) {
3228 * Set output phase alignment settings appropriate for
3231 for (i = 0; i < RW_MGR_MEM_IF_READ_DQS_WIDTH; i++) {
3232 scc_mgr_set_dqs_en_phase(i, 0);
3233 #if IO_DLL_CHAIN_LENGTH == 6
3234 scc_mgr_set_dqdqs_output_phase(i, 6);
3236 scc_mgr_set_dqdqs_output_phase(i, 7);
3241 * Write data arrives to the I/O two cycles before write
3242 * latency is reached (720 deg).
3243 * -> due to bit-slip in a/c bus
3244 * -> to allow board skew where dqs is longer than ck
3245 * -> how often can this happen!?
3246 * -> can claim back some ptaps for high freq
3247 * support if we can relax this, but i digress...
3249 * The write_clk leads mem_ck by 90 deg
3250 * The minimum ptap of the OPA is 180 deg
3251 * Each ptap has (360 / IO_DLL_CHAIN_LENGH) deg of delay
3252 * The write_clk is always delayed by 2 ptaps
3254 * Hence, to make DQS aligned to CK, we need to delay
3256 * (720 - 90 - 180 - 2 * (360 / IO_DLL_CHAIN_LENGTH))
3258 * Dividing the above by (360 / IO_DLL_CHAIN_LENGTH)
3259 * gives us the number of ptaps, which simplies to:
3261 * (1.25 * IO_DLL_CHAIN_LENGTH - 2)
3263 scc_mgr_set_dqdqs_output_phase(i, (1.25 *
3264 IO_DLL_CHAIN_LENGTH - 2));
3266 writel(0xff, &sdr_scc_mgr->dqs_ena);
3267 writel(0xff, &sdr_scc_mgr->dqs_io_ena);
3269 for (i = 0; i < RW_MGR_MEM_IF_WRITE_DQS_WIDTH; i++) {
3270 writel(i, SDR_PHYGRP_SCCGRP_ADDRESS |
3271 SCC_MGR_GROUP_COUNTER_OFFSET);
3273 writel(0xff, &sdr_scc_mgr->dq_ena);
3274 writel(0xff, &sdr_scc_mgr->dm_ena);
3275 writel(0, &sdr_scc_mgr->update);
3278 /* Compensate for simulation model behaviour */
3279 for (i = 0; i < RW_MGR_MEM_IF_READ_DQS_WIDTH; i++) {
3280 scc_mgr_set_dqs_bus_in_delay(i, 10);
3281 scc_mgr_load_dqs(i);
3283 writel(0, &sdr_scc_mgr->update);
3286 * ArriaV has hard FIFOs that can only be initialized by incrementing
3289 vfifo_offset = CALIB_VFIFO_OFFSET;
3290 for (j = 0; j < vfifo_offset; j++) {
3291 writel(0xff, &phy_mgr_cmd->inc_vfifo_hard_phy);
3293 writel(0, &phy_mgr_cmd->fifo_reset);
3296 * For ACV with hard lfifo, we get the skip-cal setting from
3297 * generation-time constant.
3299 gbl->curr_read_lat = CALIB_LFIFO_OFFSET;
3300 writel(gbl->curr_read_lat, &phy_mgr_cfg->phy_rlat);
3303 /* Memory calibration entry point */
3304 static uint32_t mem_calibrate(void)
3307 uint32_t rank_bgn, sr;
3308 uint32_t write_group, write_test_bgn;
3309 uint32_t read_group, read_test_bgn;
3310 uint32_t run_groups, current_run;
3311 uint32_t failing_groups = 0;
3312 uint32_t group_failed = 0;
3313 uint32_t sr_failed = 0;
3315 debug("%s:%d\n", __func__, __LINE__);
3316 /* Initialize the data settings */
3318 gbl->error_substage = CAL_SUBSTAGE_NIL;
3319 gbl->error_stage = CAL_STAGE_NIL;
3320 gbl->error_group = 0xff;
3326 for (i = 0; i < RW_MGR_MEM_IF_READ_DQS_WIDTH; i++) {
3327 writel(i, SDR_PHYGRP_SCCGRP_ADDRESS |
3328 SCC_MGR_GROUP_COUNTER_OFFSET);
3329 scc_set_bypass_mode(i);
3332 if ((dyn_calib_steps & CALIB_SKIP_ALL) == CALIB_SKIP_ALL) {
3334 * Set VFIFO and LFIFO to instant-on settings in skip
3337 mem_skip_calibrate();
3339 for (i = 0; i < NUM_CALIB_REPEAT; i++) {
3341 * Zero all delay chain/phase settings for all
3342 * groups and all shadow register sets.
3346 run_groups = ~param->skip_groups;
3348 for (write_group = 0, write_test_bgn = 0; write_group
3349 < RW_MGR_MEM_IF_WRITE_DQS_WIDTH; write_group++,
3350 write_test_bgn += RW_MGR_MEM_DQ_PER_WRITE_DQS) {
3351 /* Initialized the group failure */
3354 current_run = run_groups & ((1 <<
3355 RW_MGR_NUM_DQS_PER_WRITE_GROUP) - 1);
3356 run_groups = run_groups >>
3357 RW_MGR_NUM_DQS_PER_WRITE_GROUP;
3359 if (current_run == 0)
3362 writel(write_group, SDR_PHYGRP_SCCGRP_ADDRESS |
3363 SCC_MGR_GROUP_COUNTER_OFFSET);
3364 scc_mgr_zero_group(write_group, write_test_bgn,
3367 for (read_group = write_group *
3368 RW_MGR_MEM_IF_READ_DQS_WIDTH /
3369 RW_MGR_MEM_IF_WRITE_DQS_WIDTH,
3371 read_group < (write_group + 1) *
3372 RW_MGR_MEM_IF_READ_DQS_WIDTH /
3373 RW_MGR_MEM_IF_WRITE_DQS_WIDTH &&
3375 read_group++, read_test_bgn +=
3376 RW_MGR_MEM_DQ_PER_READ_DQS) {
3377 /* Calibrate the VFIFO */
3378 if (!((STATIC_CALIB_STEPS) &
3379 CALIB_SKIP_VFIFO)) {
3380 if (!rw_mgr_mem_calibrate_vfifo
3386 phy_debug_mode_flags &
3387 PHY_DEBUG_SWEEP_ALL_GROUPS)) {
3394 /* Calibrate the output side */
3395 if (group_failed == 0) {
3396 for (rank_bgn = 0, sr = 0; rank_bgn
3397 < RW_MGR_MEM_NUMBER_OF_RANKS;
3399 NUM_RANKS_PER_SHADOW_REG,
3402 if (!((STATIC_CALIB_STEPS) &
3403 CALIB_SKIP_WRITES)) {
3404 if ((STATIC_CALIB_STEPS)
3405 & CALIB_SKIP_DELAY_SWEEPS) {
3406 /* not needed in quick mode! */
3409 * Determine if this set of
3410 * ranks should be skipped
3413 if (!param->skip_shadow_regs[sr]) {
3414 if (!rw_mgr_mem_calibrate_writes
3415 (rank_bgn, write_group,
3419 phy_debug_mode_flags &
3420 PHY_DEBUG_SWEEP_ALL_GROUPS)) {
3432 if (group_failed == 0) {
3433 for (read_group = write_group *
3434 RW_MGR_MEM_IF_READ_DQS_WIDTH /
3435 RW_MGR_MEM_IF_WRITE_DQS_WIDTH,
3437 read_group < (write_group + 1)
3438 * RW_MGR_MEM_IF_READ_DQS_WIDTH
3439 / RW_MGR_MEM_IF_WRITE_DQS_WIDTH &&
3441 read_group++, read_test_bgn +=
3442 RW_MGR_MEM_DQ_PER_READ_DQS) {
3443 if (!((STATIC_CALIB_STEPS) &
3444 CALIB_SKIP_WRITES)) {
3445 if (!rw_mgr_mem_calibrate_vfifo_end
3446 (read_group, read_test_bgn)) {
3449 if (!(gbl->phy_debug_mode_flags
3450 & PHY_DEBUG_SWEEP_ALL_GROUPS)) {
3458 if (group_failed != 0)
3463 * USER If there are any failing groups then report
3466 if (failing_groups != 0)
3469 /* Calibrate the LFIFO */
3470 if (!((STATIC_CALIB_STEPS) & CALIB_SKIP_LFIFO)) {
3472 * If we're skipping groups as part of debug,
3473 * don't calibrate LFIFO.
3475 if (param->skip_groups == 0) {
3476 if (!rw_mgr_mem_calibrate_lfifo())
3484 * Do not remove this line as it makes sure all of our decisions
3485 * have been applied.
3487 writel(0, &sdr_scc_mgr->update);
3491 static uint32_t run_mem_calibrate(void)
3494 uint32_t debug_info;
3496 debug("%s:%d\n", __func__, __LINE__);
3498 /* Reset pass/fail status shown on afi_cal_success/fail */
3499 writel(PHY_MGR_CAL_RESET, &phy_mgr_cfg->cal_status);
3501 /* stop tracking manger */
3502 uint32_t ctrlcfg = readl(&sdr_ctrl->ctrl_cfg);
3504 writel(ctrlcfg & 0xFFBFFFFF, &sdr_ctrl->ctrl_cfg);
3507 rw_mgr_mem_initialize();
3509 pass = mem_calibrate();
3511 mem_precharge_and_activate();
3512 writel(0, &phy_mgr_cmd->fifo_reset);
3516 * Don't return control of the PHY back to AFI when in debug mode.
3518 if ((gbl->phy_debug_mode_flags & PHY_DEBUG_IN_DEBUG_MODE) == 0) {
3519 rw_mgr_mem_handoff();
3521 * In Hard PHY this is a 2-bit control:
3523 * 1: DDIO Mux Select
3525 writel(0x2, &phy_mgr_cfg->mux_sel);
3528 writel(ctrlcfg, &sdr_ctrl->ctrl_cfg);
3531 printf("%s: CALIBRATION PASSED\n", __FILE__);
3536 if (gbl->fom_in > 0xff)
3539 if (gbl->fom_out > 0xff)
3540 gbl->fom_out = 0xff;
3542 /* Update the FOM in the register file */
3543 debug_info = gbl->fom_in;
3544 debug_info |= gbl->fom_out << 8;
3545 writel(debug_info, &sdr_reg_file->fom);
3547 writel(debug_info, &phy_mgr_cfg->cal_debug_info);
3548 writel(PHY_MGR_CAL_SUCCESS, &phy_mgr_cfg->cal_status);
3550 printf("%s: CALIBRATION FAILED\n", __FILE__);
3552 debug_info = gbl->error_stage;
3553 debug_info |= gbl->error_substage << 8;
3554 debug_info |= gbl->error_group << 16;
3556 writel(debug_info, &sdr_reg_file->failing_stage);
3557 writel(debug_info, &phy_mgr_cfg->cal_debug_info);
3558 writel(PHY_MGR_CAL_FAIL, &phy_mgr_cfg->cal_status);
3560 /* Update the failing group/stage in the register file */
3561 debug_info = gbl->error_stage;
3562 debug_info |= gbl->error_substage << 8;
3563 debug_info |= gbl->error_group << 16;
3564 writel(debug_info, &sdr_reg_file->failing_stage);
3571 * hc_initialize_rom_data() - Initialize ROM data
3573 * Initialize ROM data.
3575 static void hc_initialize_rom_data(void)
3579 addr = SDR_PHYGRP_RWMGRGRP_ADDRESS | RW_MGR_INST_ROM_WRITE_OFFSET;
3580 for (i = 0; i < ARRAY_SIZE(inst_rom_init); i++)
3581 writel(inst_rom_init[i], addr + (i << 2));
3583 addr = SDR_PHYGRP_RWMGRGRP_ADDRESS | RW_MGR_AC_ROM_WRITE_OFFSET;
3584 for (i = 0; i < ARRAY_SIZE(ac_rom_init); i++)
3585 writel(ac_rom_init[i], addr + (i << 2));
3589 * initialize_reg_file() - Initialize SDR register file
3591 * Initialize SDR register file.
3593 static void initialize_reg_file(void)
3595 /* Initialize the register file with the correct data */
3596 writel(REG_FILE_INIT_SEQ_SIGNATURE, &sdr_reg_file->signature);
3597 writel(0, &sdr_reg_file->debug_data_addr);
3598 writel(0, &sdr_reg_file->cur_stage);
3599 writel(0, &sdr_reg_file->fom);
3600 writel(0, &sdr_reg_file->failing_stage);
3601 writel(0, &sdr_reg_file->debug1);
3602 writel(0, &sdr_reg_file->debug2);
3606 * initialize_hps_phy() - Initialize HPS PHY
3608 * Initialize HPS PHY.
3610 static void initialize_hps_phy(void)
3614 * Tracking also gets configured here because it's in the
3617 uint32_t trk_sample_count = 7500;
3618 uint32_t trk_long_idle_sample_count = (10 << 16) | 100;
3620 * Format is number of outer loops in the 16 MSB, sample
3625 reg |= SDR_CTRLGRP_PHYCTRL_PHYCTRL_0_ACDELAYEN_SET(2);
3626 reg |= SDR_CTRLGRP_PHYCTRL_PHYCTRL_0_DQDELAYEN_SET(1);
3627 reg |= SDR_CTRLGRP_PHYCTRL_PHYCTRL_0_DQSDELAYEN_SET(1);
3628 reg |= SDR_CTRLGRP_PHYCTRL_PHYCTRL_0_DQSLOGICDELAYEN_SET(1);
3629 reg |= SDR_CTRLGRP_PHYCTRL_PHYCTRL_0_RESETDELAYEN_SET(0);
3630 reg |= SDR_CTRLGRP_PHYCTRL_PHYCTRL_0_LPDDRDIS_SET(1);
3632 * This field selects the intrinsic latency to RDATA_EN/FULL path.
3633 * 00-bypass, 01- add 5 cycles, 10- add 10 cycles, 11- add 15 cycles.
3635 reg |= SDR_CTRLGRP_PHYCTRL_PHYCTRL_0_ADDLATSEL_SET(0);
3636 reg |= SDR_CTRLGRP_PHYCTRL_PHYCTRL_0_SAMPLECOUNT_19_0_SET(
3638 writel(reg, &sdr_ctrl->phy_ctrl0);
3641 reg |= SDR_CTRLGRP_PHYCTRL_PHYCTRL_1_SAMPLECOUNT_31_20_SET(
3643 SDR_CTRLGRP_PHYCTRL_PHYCTRL_0_SAMPLECOUNT_19_0_WIDTH);
3644 reg |= SDR_CTRLGRP_PHYCTRL_PHYCTRL_1_LONGIDLESAMPLECOUNT_19_0_SET(
3645 trk_long_idle_sample_count);
3646 writel(reg, &sdr_ctrl->phy_ctrl1);
3649 reg |= SDR_CTRLGRP_PHYCTRL_PHYCTRL_2_LONGIDLESAMPLECOUNT_31_20_SET(
3650 trk_long_idle_sample_count >>
3651 SDR_CTRLGRP_PHYCTRL_PHYCTRL_1_LONGIDLESAMPLECOUNT_19_0_WIDTH);
3652 writel(reg, &sdr_ctrl->phy_ctrl2);
3655 static void initialize_tracking(void)
3657 uint32_t concatenated_longidle = 0x0;
3658 uint32_t concatenated_delays = 0x0;
3659 uint32_t concatenated_rw_addr = 0x0;
3660 uint32_t concatenated_refresh = 0x0;
3661 uint32_t trk_sample_count = 7500;
3662 uint32_t dtaps_per_ptap;
3666 * compute usable version of value in case we skip full
3671 while (tmp_delay < IO_DELAY_PER_OPA_TAP) {
3673 tmp_delay += IO_DELAY_PER_DCHAIN_TAP;
3677 concatenated_longidle = concatenated_longidle ^ 10;
3678 /*longidle outer loop */
3679 concatenated_longidle = concatenated_longidle << 16;
3680 concatenated_longidle = concatenated_longidle ^ 100;
3681 /*longidle sample count */
3682 concatenated_delays = concatenated_delays ^ 243;
3683 /* trfc, worst case of 933Mhz 4Gb */
3684 concatenated_delays = concatenated_delays << 8;
3685 concatenated_delays = concatenated_delays ^ 14;
3686 /* trcd, worst case */
3687 concatenated_delays = concatenated_delays << 8;
3688 concatenated_delays = concatenated_delays ^ 10;
3690 concatenated_delays = concatenated_delays << 8;
3691 concatenated_delays = concatenated_delays ^ 4;
3694 concatenated_rw_addr = concatenated_rw_addr ^ RW_MGR_IDLE;
3695 concatenated_rw_addr = concatenated_rw_addr << 8;
3696 concatenated_rw_addr = concatenated_rw_addr ^ RW_MGR_ACTIVATE_1;
3697 concatenated_rw_addr = concatenated_rw_addr << 8;
3698 concatenated_rw_addr = concatenated_rw_addr ^ RW_MGR_SGLE_READ;
3699 concatenated_rw_addr = concatenated_rw_addr << 8;
3700 concatenated_rw_addr = concatenated_rw_addr ^ RW_MGR_PRECHARGE_ALL;
3702 concatenated_refresh = concatenated_refresh ^ RW_MGR_REFRESH_ALL;
3703 concatenated_refresh = concatenated_refresh << 24;
3704 concatenated_refresh = concatenated_refresh ^ 1000; /* trefi */
3706 /* Initialize the register file with the correct data */
3707 writel(dtaps_per_ptap, &sdr_reg_file->dtaps_per_ptap);
3708 writel(trk_sample_count, &sdr_reg_file->trk_sample_count);
3709 writel(concatenated_longidle, &sdr_reg_file->trk_longidle);
3710 writel(concatenated_delays, &sdr_reg_file->delays);
3711 writel(concatenated_rw_addr, &sdr_reg_file->trk_rw_mgr_addr);
3712 writel(RW_MGR_MEM_IF_READ_DQS_WIDTH, &sdr_reg_file->trk_read_dqs_width);
3713 writel(concatenated_refresh, &sdr_reg_file->trk_rfsh);
3716 int sdram_calibration_full(void)
3718 struct param_type my_param;
3719 struct gbl_type my_gbl;
3726 /* Initialize the debug mode flags */
3727 gbl->phy_debug_mode_flags = 0;
3728 /* Set the calibration enabled by default */
3729 gbl->phy_debug_mode_flags |= PHY_DEBUG_ENABLE_CAL_RPT;
3731 * Only sweep all groups (regardless of fail state) by default
3732 * Set enabled read test by default.
3734 #if DISABLE_GUARANTEED_READ
3735 gbl->phy_debug_mode_flags |= PHY_DEBUG_DISABLE_GUARANTEED_READ;
3737 /* Initialize the register file */
3738 initialize_reg_file();
3740 /* Initialize any PHY CSR */
3741 initialize_hps_phy();
3743 scc_mgr_initialize();
3745 initialize_tracking();
3747 /* USER Enable all ranks, groups */
3748 for (i = 0; i < RW_MGR_MEM_NUMBER_OF_RANKS; i++)
3749 param->skip_ranks[i] = 0;
3750 for (i = 0; i < NUM_SHADOW_REGS; ++i)
3751 param->skip_shadow_regs[i] = 0;
3752 param->skip_groups = 0;
3754 printf("%s: Preparing to start memory calibration\n", __FILE__);
3756 debug("%s:%d\n", __func__, __LINE__);
3757 debug_cond(DLEVEL == 1,
3758 "DDR3 FULL_RATE ranks=%u cs/dimm=%u dq/dqs=%u,%u vg/dqs=%u,%u ",
3759 RW_MGR_MEM_NUMBER_OF_RANKS, RW_MGR_MEM_NUMBER_OF_CS_PER_DIMM,
3760 RW_MGR_MEM_DQ_PER_READ_DQS, RW_MGR_MEM_DQ_PER_WRITE_DQS,
3761 RW_MGR_MEM_VIRTUAL_GROUPS_PER_READ_DQS,
3762 RW_MGR_MEM_VIRTUAL_GROUPS_PER_WRITE_DQS);
3763 debug_cond(DLEVEL == 1,
3764 "dqs=%u,%u dq=%u dm=%u ptap_delay=%u dtap_delay=%u ",
3765 RW_MGR_MEM_IF_READ_DQS_WIDTH, RW_MGR_MEM_IF_WRITE_DQS_WIDTH,
3766 RW_MGR_MEM_DATA_WIDTH, RW_MGR_MEM_DATA_MASK_WIDTH,
3767 IO_DELAY_PER_OPA_TAP, IO_DELAY_PER_DCHAIN_TAP);
3768 debug_cond(DLEVEL == 1, "dtap_dqsen_delay=%u, dll=%u",
3769 IO_DELAY_PER_DQS_EN_DCHAIN_TAP, IO_DLL_CHAIN_LENGTH);
3770 debug_cond(DLEVEL == 1, "max values: en_p=%u dqdqs_p=%u en_d=%u dqs_in_d=%u ",
3771 IO_DQS_EN_PHASE_MAX, IO_DQDQS_OUT_PHASE_MAX,
3772 IO_DQS_EN_DELAY_MAX, IO_DQS_IN_DELAY_MAX);
3773 debug_cond(DLEVEL == 1, "io_in_d=%u io_out1_d=%u io_out2_d=%u ",
3774 IO_IO_IN_DELAY_MAX, IO_IO_OUT1_DELAY_MAX,
3775 IO_IO_OUT2_DELAY_MAX);
3776 debug_cond(DLEVEL == 1, "dqs_in_reserve=%u dqs_out_reserve=%u\n",
3777 IO_DQS_IN_RESERVE, IO_DQS_OUT_RESERVE);
3779 hc_initialize_rom_data();
3781 /* update info for sims */
3782 reg_file_set_stage(CAL_STAGE_NIL);
3783 reg_file_set_group(0);
3786 * Load global needed for those actions that require
3787 * some dynamic calibration support.
3789 dyn_calib_steps = STATIC_CALIB_STEPS;
3791 * Load global to allow dynamic selection of delay loop settings
3792 * based on calibration mode.
3794 if (!(dyn_calib_steps & CALIB_SKIP_DELAY_LOOPS))
3795 skip_delay_mask = 0xff;
3797 skip_delay_mask = 0x0;
3799 pass = run_mem_calibrate();
3801 printf("%s: Calibration complete\n", __FILE__);