2 * Copyright Altera Corporation (C) 2012-2015
4 * SPDX-License-Identifier: BSD-3-Clause
9 #include <asm/arch/sdram.h>
11 #include "sequencer.h"
12 #include "sequencer_auto.h"
13 #include "sequencer_auto_ac_init.h"
14 #include "sequencer_auto_inst_init.h"
15 #include "sequencer_defines.h"
17 static struct socfpga_sdr_rw_load_manager *sdr_rw_load_mgr_regs =
18 (struct socfpga_sdr_rw_load_manager *)(SDR_PHYGRP_RWMGRGRP_ADDRESS | 0x800);
20 static struct socfpga_sdr_rw_load_jump_manager *sdr_rw_load_jump_mgr_regs =
21 (struct socfpga_sdr_rw_load_jump_manager *)(SDR_PHYGRP_RWMGRGRP_ADDRESS | 0xC00);
23 static struct socfpga_sdr_reg_file *sdr_reg_file =
24 (struct socfpga_sdr_reg_file *)SDR_PHYGRP_REGFILEGRP_ADDRESS;
26 static struct socfpga_sdr_scc_mgr *sdr_scc_mgr =
27 (struct socfpga_sdr_scc_mgr *)(SDR_PHYGRP_SCCGRP_ADDRESS | 0xe00);
29 static struct socfpga_phy_mgr_cmd *phy_mgr_cmd =
30 (struct socfpga_phy_mgr_cmd *)SDR_PHYGRP_PHYMGRGRP_ADDRESS;
32 static struct socfpga_phy_mgr_cfg *phy_mgr_cfg =
33 (struct socfpga_phy_mgr_cfg *)(SDR_PHYGRP_PHYMGRGRP_ADDRESS | 0x40);
35 static struct socfpga_data_mgr *data_mgr =
36 (struct socfpga_data_mgr *)SDR_PHYGRP_DATAMGRGRP_ADDRESS;
38 static struct socfpga_sdr_ctrl *sdr_ctrl =
39 (struct socfpga_sdr_ctrl *)SDR_CTRLGRP_ADDRESS;
44 * In order to reduce ROM size, most of the selectable calibration steps are
45 * decided at compile time based on the user's calibration mode selection,
46 * as captured by the STATIC_CALIB_STEPS selection below.
48 * However, to support simulation-time selection of fast simulation mode, where
49 * we skip everything except the bare minimum, we need a few of the steps to
50 * be dynamic. In those cases, we either use the DYNAMIC_CALIB_STEPS for the
51 * check, which is based on the rtl-supplied value, or we dynamically compute
52 * the value to use based on the dynamically-chosen calibration mode
56 #define STATIC_IN_RTL_SIM 0
57 #define STATIC_SKIP_DELAY_LOOPS 0
59 #define STATIC_CALIB_STEPS (STATIC_IN_RTL_SIM | CALIB_SKIP_FULL_TEST | \
60 STATIC_SKIP_DELAY_LOOPS)
62 /* calibration steps requested by the rtl */
63 uint16_t dyn_calib_steps;
66 * To make CALIB_SKIP_DELAY_LOOPS a dynamic conditional option
67 * instead of static, we use boolean logic to select between
68 * non-skip and skip values
70 * The mask is set to include all bits when not-skipping, but is
74 uint16_t skip_delay_mask; /* mask off bits when skipping/not-skipping */
76 #define SKIP_DELAY_LOOP_VALUE_OR_ZERO(non_skip_value) \
77 ((non_skip_value) & skip_delay_mask)
80 struct param_type *param;
81 uint32_t curr_shadow_reg;
83 static uint32_t rw_mgr_mem_calibrate_write_test(uint32_t rank_bgn,
84 uint32_t write_group, uint32_t use_dm,
85 uint32_t all_correct, uint32_t *bit_chk, uint32_t all_ranks);
87 static void set_failing_group_stage(uint32_t group, uint32_t stage,
91 * Only set the global stage if there was not been any other
94 if (gbl->error_stage == CAL_STAGE_NIL) {
95 gbl->error_substage = substage;
96 gbl->error_stage = stage;
97 gbl->error_group = group;
101 static void reg_file_set_group(u16 set_group)
103 clrsetbits_le32(&sdr_reg_file->cur_stage, 0xffff0000, set_group << 16);
106 static void reg_file_set_stage(u8 set_stage)
108 clrsetbits_le32(&sdr_reg_file->cur_stage, 0xffff, set_stage & 0xff);
111 static void reg_file_set_sub_stage(u8 set_sub_stage)
113 set_sub_stage &= 0xff;
114 clrsetbits_le32(&sdr_reg_file->cur_stage, 0xff00, set_sub_stage << 8);
118 * phy_mgr_initialize() - Initialize PHY Manager
120 * Initialize PHY Manager.
122 static void phy_mgr_initialize(void)
126 debug("%s:%d\n", __func__, __LINE__);
127 /* Calibration has control over path to memory */
129 * In Hard PHY this is a 2-bit control:
133 writel(0x3, &phy_mgr_cfg->mux_sel);
135 /* USER memory clock is not stable we begin initialization */
136 writel(0, &phy_mgr_cfg->reset_mem_stbl);
138 /* USER calibration status all set to zero */
139 writel(0, &phy_mgr_cfg->cal_status);
141 writel(0, &phy_mgr_cfg->cal_debug_info);
143 /* Init params only if we do NOT skip calibration. */
144 if ((dyn_calib_steps & CALIB_SKIP_ALL) == CALIB_SKIP_ALL)
147 ratio = RW_MGR_MEM_DQ_PER_READ_DQS /
148 RW_MGR_MEM_VIRTUAL_GROUPS_PER_READ_DQS;
149 param->read_correct_mask_vg = (1 << ratio) - 1;
150 param->write_correct_mask_vg = (1 << ratio) - 1;
151 param->read_correct_mask = (1 << RW_MGR_MEM_DQ_PER_READ_DQS) - 1;
152 param->write_correct_mask = (1 << RW_MGR_MEM_DQ_PER_WRITE_DQS) - 1;
153 ratio = RW_MGR_MEM_DATA_WIDTH /
154 RW_MGR_MEM_DATA_MASK_WIDTH;
155 param->dm_correct_mask = (1 << ratio) - 1;
159 * set_rank_and_odt_mask() - Set Rank and ODT mask
161 * @odt_mode: ODT mode, OFF or READ_WRITE
163 * Set Rank and ODT mask (On-Die Termination).
165 static void set_rank_and_odt_mask(const u32 rank, const u32 odt_mode)
171 if (odt_mode == RW_MGR_ODT_MODE_OFF) {
174 } else { /* RW_MGR_ODT_MODE_READ_WRITE */
175 switch (RW_MGR_MEM_NUMBER_OF_RANKS) {
177 /* Read: ODT = 0 ; Write: ODT = 1 */
181 case 2: /* 2 Ranks */
182 if (RW_MGR_MEM_NUMBER_OF_CS_PER_DIMM == 1) {
184 * - Dual-Slot , Single-Rank (1 CS per DIMM)
186 * - RDIMM, 4 total CS (2 CS per DIMM, 2 DIMM)
188 * Since MEM_NUMBER_OF_RANKS is 2, they
189 * are both single rank with 2 CS each
190 * (special for RDIMM).
192 * Read: Turn on ODT on the opposite rank
193 * Write: Turn on ODT on all ranks
195 odt_mask_0 = 0x3 & ~(1 << rank);
199 * - Single-Slot , Dual-Rank (2 CS per DIMM)
201 * Read: Turn on ODT off on all ranks
202 * Write: Turn on ODT on active rank
205 odt_mask_1 = 0x3 & (1 << rank);
208 case 4: /* 4 Ranks */
210 * ----------+-----------------------+
212 * Read From +-----------------------+
213 * Rank | 3 | 2 | 1 | 0 |
214 * ----------+-----+-----+-----+-----+
215 * 0 | 0 | 1 | 0 | 0 |
216 * 1 | 1 | 0 | 0 | 0 |
217 * 2 | 0 | 0 | 0 | 1 |
218 * 3 | 0 | 0 | 1 | 0 |
219 * ----------+-----+-----+-----+-----+
222 * ----------+-----------------------+
224 * Write To +-----------------------+
225 * Rank | 3 | 2 | 1 | 0 |
226 * ----------+-----+-----+-----+-----+
227 * 0 | 0 | 1 | 0 | 1 |
228 * 1 | 1 | 0 | 1 | 0 |
229 * 2 | 0 | 1 | 0 | 1 |
230 * 3 | 1 | 0 | 1 | 0 |
231 * ----------+-----+-----+-----+-----+
255 cs_and_odt_mask = (0xFF & ~(1 << rank)) |
256 ((0xFF & odt_mask_0) << 8) |
257 ((0xFF & odt_mask_1) << 16);
258 writel(cs_and_odt_mask, SDR_PHYGRP_RWMGRGRP_ADDRESS |
259 RW_MGR_SET_CS_AND_ODT_MASK_OFFSET);
263 * scc_mgr_set() - Set SCC Manager register
264 * @off: Base offset in SCC Manager space
265 * @grp: Read/Write group
266 * @val: Value to be set
268 * This function sets the SCC Manager (Scan Chain Control Manager) register.
270 static void scc_mgr_set(u32 off, u32 grp, u32 val)
272 writel(val, SDR_PHYGRP_SCCGRP_ADDRESS | off | (grp << 2));
276 * scc_mgr_initialize() - Initialize SCC Manager registers
278 * Initialize SCC Manager registers.
280 static void scc_mgr_initialize(void)
283 * Clear register file for HPS. 16 (2^4) is the size of the
284 * full register file in the scc mgr:
285 * RFILE_DEPTH = 1 + log2(MEM_DQ_PER_DQS + 1 + MEM_DM_PER_DQS +
286 * MEM_IF_READ_DQS_WIDTH - 1);
290 for (i = 0; i < 16; i++) {
291 debug_cond(DLEVEL == 1, "%s:%d: Clearing SCC RFILE index %u\n",
292 __func__, __LINE__, i);
293 scc_mgr_set(SCC_MGR_HHP_RFILE_OFFSET, 0, i);
297 static void scc_mgr_set_dqdqs_output_phase(uint32_t write_group, uint32_t phase)
299 scc_mgr_set(SCC_MGR_DQDQS_OUT_PHASE_OFFSET, write_group, phase);
302 static void scc_mgr_set_dqs_bus_in_delay(uint32_t read_group, uint32_t delay)
304 scc_mgr_set(SCC_MGR_DQS_IN_DELAY_OFFSET, read_group, delay);
307 static void scc_mgr_set_dqs_en_phase(uint32_t read_group, uint32_t phase)
309 scc_mgr_set(SCC_MGR_DQS_EN_PHASE_OFFSET, read_group, phase);
312 static void scc_mgr_set_dqs_en_delay(uint32_t read_group, uint32_t delay)
314 scc_mgr_set(SCC_MGR_DQS_EN_DELAY_OFFSET, read_group, delay);
317 static void scc_mgr_set_dqs_io_in_delay(uint32_t delay)
319 scc_mgr_set(SCC_MGR_IO_IN_DELAY_OFFSET, RW_MGR_MEM_DQ_PER_WRITE_DQS,
323 static void scc_mgr_set_dq_in_delay(uint32_t dq_in_group, uint32_t delay)
325 scc_mgr_set(SCC_MGR_IO_IN_DELAY_OFFSET, dq_in_group, delay);
328 static void scc_mgr_set_dq_out1_delay(uint32_t dq_in_group, uint32_t delay)
330 scc_mgr_set(SCC_MGR_IO_OUT1_DELAY_OFFSET, dq_in_group, delay);
333 static void scc_mgr_set_dqs_out1_delay(uint32_t delay)
335 scc_mgr_set(SCC_MGR_IO_OUT1_DELAY_OFFSET, RW_MGR_MEM_DQ_PER_WRITE_DQS,
339 static void scc_mgr_set_dm_out1_delay(uint32_t dm, uint32_t delay)
341 scc_mgr_set(SCC_MGR_IO_OUT1_DELAY_OFFSET,
342 RW_MGR_MEM_DQ_PER_WRITE_DQS + 1 + dm,
346 /* load up dqs config settings */
347 static void scc_mgr_load_dqs(uint32_t dqs)
349 writel(dqs, &sdr_scc_mgr->dqs_ena);
352 /* load up dqs io config settings */
353 static void scc_mgr_load_dqs_io(void)
355 writel(0, &sdr_scc_mgr->dqs_io_ena);
358 /* load up dq config settings */
359 static void scc_mgr_load_dq(uint32_t dq_in_group)
361 writel(dq_in_group, &sdr_scc_mgr->dq_ena);
364 /* load up dm config settings */
365 static void scc_mgr_load_dm(uint32_t dm)
367 writel(dm, &sdr_scc_mgr->dm_ena);
371 * scc_mgr_set_all_ranks() - Set SCC Manager register for all ranks
372 * @off: Base offset in SCC Manager space
373 * @grp: Read/Write group
374 * @val: Value to be set
375 * @update: If non-zero, trigger SCC Manager update for all ranks
377 * This function sets the SCC Manager (Scan Chain Control Manager) register
378 * and optionally triggers the SCC update for all ranks.
380 static void scc_mgr_set_all_ranks(const u32 off, const u32 grp, const u32 val,
385 for (r = 0; r < RW_MGR_MEM_NUMBER_OF_RANKS;
386 r += NUM_RANKS_PER_SHADOW_REG) {
387 scc_mgr_set(off, grp, val);
389 if (update || (r == 0)) {
390 writel(grp, &sdr_scc_mgr->dqs_ena);
391 writel(0, &sdr_scc_mgr->update);
396 static void scc_mgr_set_dqs_en_phase_all_ranks(u32 read_group, u32 phase)
399 * USER although the h/w doesn't support different phases per
400 * shadow register, for simplicity our scc manager modeling
401 * keeps different phase settings per shadow reg, and it's
402 * important for us to keep them in sync to match h/w.
403 * for efficiency, the scan chain update should occur only
406 scc_mgr_set_all_ranks(SCC_MGR_DQS_EN_PHASE_OFFSET,
407 read_group, phase, 0);
410 static void scc_mgr_set_dqdqs_output_phase_all_ranks(uint32_t write_group,
414 * USER although the h/w doesn't support different phases per
415 * shadow register, for simplicity our scc manager modeling
416 * keeps different phase settings per shadow reg, and it's
417 * important for us to keep them in sync to match h/w.
418 * for efficiency, the scan chain update should occur only
421 scc_mgr_set_all_ranks(SCC_MGR_DQDQS_OUT_PHASE_OFFSET,
422 write_group, phase, 0);
425 static void scc_mgr_set_dqs_en_delay_all_ranks(uint32_t read_group,
429 * In shadow register mode, the T11 settings are stored in
430 * registers in the core, which are updated by the DQS_ENA
431 * signals. Not issuing the SCC_MGR_UPD command allows us to
432 * save lots of rank switching overhead, by calling
433 * select_shadow_regs_for_update with update_scan_chains
436 scc_mgr_set_all_ranks(SCC_MGR_DQS_EN_DELAY_OFFSET,
437 read_group, delay, 1);
438 writel(0, &sdr_scc_mgr->update);
442 * scc_mgr_set_oct_out1_delay() - Set OCT output delay
443 * @write_group: Write group
444 * @delay: Delay value
446 * This function sets the OCT output delay in SCC manager.
448 static void scc_mgr_set_oct_out1_delay(const u32 write_group, const u32 delay)
450 const int ratio = RW_MGR_MEM_IF_READ_DQS_WIDTH /
451 RW_MGR_MEM_IF_WRITE_DQS_WIDTH;
452 const int base = write_group * ratio;
455 * Load the setting in the SCC manager
456 * Although OCT affects only write data, the OCT delay is controlled
457 * by the DQS logic block which is instantiated once per read group.
458 * For protocols where a write group consists of multiple read groups,
459 * the setting must be set multiple times.
461 for (i = 0; i < ratio; i++)
462 scc_mgr_set(SCC_MGR_OCT_OUT1_DELAY_OFFSET, base + i, delay);
466 * scc_mgr_set_hhp_extras() - Set HHP extras.
468 * Load the fixed setting in the SCC manager HHP extras.
470 static void scc_mgr_set_hhp_extras(void)
473 * Load the fixed setting in the SCC manager
474 * bits: 0:0 = 1'b1 - DQS bypass
475 * bits: 1:1 = 1'b1 - DQ bypass
476 * bits: 4:2 = 3'b001 - rfifo_mode
477 * bits: 6:5 = 2'b01 - rfifo clock_select
478 * bits: 7:7 = 1'b0 - separate gating from ungating setting
479 * bits: 8:8 = 1'b0 - separate OE from Output delay setting
481 const u32 value = (0 << 8) | (0 << 7) | (1 << 5) |
482 (1 << 2) | (1 << 1) | (1 << 0);
483 const u32 addr = SDR_PHYGRP_SCCGRP_ADDRESS |
484 SCC_MGR_HHP_GLOBALS_OFFSET |
485 SCC_MGR_HHP_EXTRAS_OFFSET;
487 debug_cond(DLEVEL == 1, "%s:%d Setting HHP Extras\n",
490 debug_cond(DLEVEL == 1, "%s:%d Done Setting HHP Extras\n",
495 * scc_mgr_zero_all() - Zero all DQS config
497 * Zero all DQS config.
499 static void scc_mgr_zero_all(void)
504 * USER Zero all DQS config settings, across all groups and all
507 for (r = 0; r < RW_MGR_MEM_NUMBER_OF_RANKS;
508 r += NUM_RANKS_PER_SHADOW_REG) {
509 for (i = 0; i < RW_MGR_MEM_IF_READ_DQS_WIDTH; i++) {
511 * The phases actually don't exist on a per-rank basis,
512 * but there's no harm updating them several times, so
513 * let's keep the code simple.
515 scc_mgr_set_dqs_bus_in_delay(i, IO_DQS_IN_RESERVE);
516 scc_mgr_set_dqs_en_phase(i, 0);
517 scc_mgr_set_dqs_en_delay(i, 0);
520 for (i = 0; i < RW_MGR_MEM_IF_WRITE_DQS_WIDTH; i++) {
521 scc_mgr_set_dqdqs_output_phase(i, 0);
522 /* Arria V/Cyclone V don't have out2. */
523 scc_mgr_set_oct_out1_delay(i, IO_DQS_OUT_RESERVE);
527 /* Multicast to all DQS group enables. */
528 writel(0xff, &sdr_scc_mgr->dqs_ena);
529 writel(0, &sdr_scc_mgr->update);
533 * scc_set_bypass_mode() - Set bypass mode and trigger SCC update
534 * @write_group: Write group
536 * Set bypass mode and trigger SCC update.
538 static void scc_set_bypass_mode(const u32 write_group)
540 /* Multicast to all DQ enables. */
541 writel(0xff, &sdr_scc_mgr->dq_ena);
542 writel(0xff, &sdr_scc_mgr->dm_ena);
544 /* Update current DQS IO enable. */
545 writel(0, &sdr_scc_mgr->dqs_io_ena);
547 /* Update the DQS logic. */
548 writel(write_group, &sdr_scc_mgr->dqs_ena);
551 writel(0, &sdr_scc_mgr->update);
555 * scc_mgr_load_dqs_for_write_group() - Load DQS settings for Write Group
556 * @write_group: Write group
558 * Load DQS settings for Write Group, do not trigger SCC update.
560 static void scc_mgr_load_dqs_for_write_group(const u32 write_group)
562 const int ratio = RW_MGR_MEM_IF_READ_DQS_WIDTH /
563 RW_MGR_MEM_IF_WRITE_DQS_WIDTH;
564 const int base = write_group * ratio;
567 * Load the setting in the SCC manager
568 * Although OCT affects only write data, the OCT delay is controlled
569 * by the DQS logic block which is instantiated once per read group.
570 * For protocols where a write group consists of multiple read groups,
571 * the setting must be set multiple times.
573 for (i = 0; i < ratio; i++)
574 writel(base + i, &sdr_scc_mgr->dqs_ena);
578 * scc_mgr_zero_group() - Zero all configs for a group
580 * Zero DQ, DM, DQS and OCT configs for a group.
582 static void scc_mgr_zero_group(const u32 write_group, const int out_only)
586 for (r = 0; r < RW_MGR_MEM_NUMBER_OF_RANKS;
587 r += NUM_RANKS_PER_SHADOW_REG) {
588 /* Zero all DQ config settings. */
589 for (i = 0; i < RW_MGR_MEM_DQ_PER_WRITE_DQS; i++) {
590 scc_mgr_set_dq_out1_delay(i, 0);
592 scc_mgr_set_dq_in_delay(i, 0);
595 /* Multicast to all DQ enables. */
596 writel(0xff, &sdr_scc_mgr->dq_ena);
598 /* Zero all DM config settings. */
599 for (i = 0; i < RW_MGR_NUM_DM_PER_WRITE_GROUP; i++)
600 scc_mgr_set_dm_out1_delay(i, 0);
602 /* Multicast to all DM enables. */
603 writel(0xff, &sdr_scc_mgr->dm_ena);
605 /* Zero all DQS IO settings. */
607 scc_mgr_set_dqs_io_in_delay(0);
609 /* Arria V/Cyclone V don't have out2. */
610 scc_mgr_set_dqs_out1_delay(IO_DQS_OUT_RESERVE);
611 scc_mgr_set_oct_out1_delay(write_group, IO_DQS_OUT_RESERVE);
612 scc_mgr_load_dqs_for_write_group(write_group);
614 /* Multicast to all DQS IO enables (only 1 in total). */
615 writel(0, &sdr_scc_mgr->dqs_io_ena);
617 /* Hit update to zero everything. */
618 writel(0, &sdr_scc_mgr->update);
623 * apply and load a particular input delay for the DQ pins in a group
624 * group_bgn is the index of the first dq pin (in the write group)
626 static void scc_mgr_apply_group_dq_in_delay(uint32_t group_bgn, uint32_t delay)
630 for (i = 0, p = group_bgn; i < RW_MGR_MEM_DQ_PER_READ_DQS; i++, p++) {
631 scc_mgr_set_dq_in_delay(p, delay);
637 * scc_mgr_apply_group_dq_out1_delay() - Apply and load an output delay for the DQ pins in a group
638 * @delay: Delay value
640 * Apply and load a particular output delay for the DQ pins in a group.
642 static void scc_mgr_apply_group_dq_out1_delay(const u32 delay)
646 for (i = 0; i < RW_MGR_MEM_DQ_PER_WRITE_DQS; i++) {
647 scc_mgr_set_dq_out1_delay(i, delay);
652 /* apply and load a particular output delay for the DM pins in a group */
653 static void scc_mgr_apply_group_dm_out1_delay(uint32_t delay1)
657 for (i = 0; i < RW_MGR_NUM_DM_PER_WRITE_GROUP; i++) {
658 scc_mgr_set_dm_out1_delay(i, delay1);
664 /* apply and load delay on both DQS and OCT out1 */
665 static void scc_mgr_apply_group_dqs_io_and_oct_out1(uint32_t write_group,
668 scc_mgr_set_dqs_out1_delay(delay);
669 scc_mgr_load_dqs_io();
671 scc_mgr_set_oct_out1_delay(write_group, delay);
672 scc_mgr_load_dqs_for_write_group(write_group);
676 * scc_mgr_apply_group_all_out_delay_add() - Apply a delay to the entire output side: DQ, DM, DQS, OCT
677 * @write_group: Write group
678 * @delay: Delay value
680 * Apply a delay to the entire output side: DQ, DM, DQS, OCT.
682 static void scc_mgr_apply_group_all_out_delay_add(const u32 write_group,
688 for (i = 0; i < RW_MGR_MEM_DQ_PER_WRITE_DQS; i++)
692 for (i = 0; i < RW_MGR_NUM_DM_PER_WRITE_GROUP; i++)
696 new_delay = READ_SCC_DQS_IO_OUT2_DELAY + delay;
697 if (new_delay > IO_IO_OUT2_DELAY_MAX) {
698 debug_cond(DLEVEL == 1,
699 "%s:%d (%u, %u) DQS: %u > %d; adding %u to OUT1\n",
700 __func__, __LINE__, write_group, delay, new_delay,
701 IO_IO_OUT2_DELAY_MAX,
702 new_delay - IO_IO_OUT2_DELAY_MAX);
703 new_delay -= IO_IO_OUT2_DELAY_MAX;
704 scc_mgr_set_dqs_out1_delay(new_delay);
707 scc_mgr_load_dqs_io();
710 new_delay = READ_SCC_OCT_OUT2_DELAY + delay;
711 if (new_delay > IO_IO_OUT2_DELAY_MAX) {
712 debug_cond(DLEVEL == 1,
713 "%s:%d (%u, %u) DQS: %u > %d; adding %u to OUT1\n",
714 __func__, __LINE__, write_group, delay,
715 new_delay, IO_IO_OUT2_DELAY_MAX,
716 new_delay - IO_IO_OUT2_DELAY_MAX);
717 new_delay -= IO_IO_OUT2_DELAY_MAX;
718 scc_mgr_set_oct_out1_delay(write_group, new_delay);
721 scc_mgr_load_dqs_for_write_group(write_group);
725 * scc_mgr_apply_group_all_out_delay_add() - Apply a delay to the entire output side to all ranks
726 * @write_group: Write group
727 * @delay: Delay value
729 * Apply a delay to the entire output side (DQ, DM, DQS, OCT) to all ranks.
732 scc_mgr_apply_group_all_out_delay_add_all_ranks(const u32 write_group,
737 for (r = 0; r < RW_MGR_MEM_NUMBER_OF_RANKS;
738 r += NUM_RANKS_PER_SHADOW_REG) {
739 scc_mgr_apply_group_all_out_delay_add(write_group, delay);
740 writel(0, &sdr_scc_mgr->update);
745 * set_jump_as_return() - Return instruction optimization
747 * Optimization used to recover some slots in ddr3 inst_rom could be
748 * applied to other protocols if we wanted to
750 static void set_jump_as_return(void)
753 * To save space, we replace return with jump to special shared
754 * RETURN instruction so we set the counter to large value so that
757 writel(0xff, &sdr_rw_load_mgr_regs->load_cntr0);
758 writel(RW_MGR_RETURN, &sdr_rw_load_jump_mgr_regs->load_jump_add0);
762 * should always use constants as argument to ensure all computations are
763 * performed at compile time
765 static void delay_for_n_mem_clocks(const uint32_t clocks)
772 debug("%s:%d: clocks=%u ... start\n", __func__, __LINE__, clocks);
775 afi_clocks = (clocks + AFI_RATE_RATIO-1) / AFI_RATE_RATIO;
776 /* scale (rounding up) to get afi clocks */
779 * Note, we don't bother accounting for being off a little bit
780 * because of a few extra instructions in outer loops
781 * Note, the loops have a test at the end, and do the test before
782 * the decrement, and so always perform the loop
783 * 1 time more than the counter value
785 if (afi_clocks == 0) {
787 } else if (afi_clocks <= 0x100) {
788 inner = afi_clocks-1;
791 } else if (afi_clocks <= 0x10000) {
793 outer = (afi_clocks-1) >> 8;
798 c_loop = (afi_clocks-1) >> 16;
802 * rom instructions are structured as follows:
804 * IDLE_LOOP2: jnz cntr0, TARGET_A
805 * IDLE_LOOP1: jnz cntr1, TARGET_B
808 * so, when doing nested loops, TARGET_A is set to IDLE_LOOP2, and
809 * TARGET_B is set to IDLE_LOOP2 as well
811 * if we have no outer loop, though, then we can use IDLE_LOOP1 only,
812 * and set TARGET_B to IDLE_LOOP1 and we skip IDLE_LOOP2 entirely
814 * a little confusing, but it helps save precious space in the inst_rom
815 * and sequencer rom and keeps the delays more accurate and reduces
818 if (afi_clocks <= 0x100) {
819 writel(SKIP_DELAY_LOOP_VALUE_OR_ZERO(inner),
820 &sdr_rw_load_mgr_regs->load_cntr1);
822 writel(RW_MGR_IDLE_LOOP1,
823 &sdr_rw_load_jump_mgr_regs->load_jump_add1);
825 writel(RW_MGR_IDLE_LOOP1, SDR_PHYGRP_RWMGRGRP_ADDRESS |
826 RW_MGR_RUN_SINGLE_GROUP_OFFSET);
828 writel(SKIP_DELAY_LOOP_VALUE_OR_ZERO(inner),
829 &sdr_rw_load_mgr_regs->load_cntr0);
831 writel(SKIP_DELAY_LOOP_VALUE_OR_ZERO(outer),
832 &sdr_rw_load_mgr_regs->load_cntr1);
834 writel(RW_MGR_IDLE_LOOP2,
835 &sdr_rw_load_jump_mgr_regs->load_jump_add0);
837 writel(RW_MGR_IDLE_LOOP2,
838 &sdr_rw_load_jump_mgr_regs->load_jump_add1);
840 /* hack to get around compiler not being smart enough */
841 if (afi_clocks <= 0x10000) {
842 /* only need to run once */
843 writel(RW_MGR_IDLE_LOOP2, SDR_PHYGRP_RWMGRGRP_ADDRESS |
844 RW_MGR_RUN_SINGLE_GROUP_OFFSET);
847 writel(RW_MGR_IDLE_LOOP2,
848 SDR_PHYGRP_RWMGRGRP_ADDRESS |
849 RW_MGR_RUN_SINGLE_GROUP_OFFSET);
850 } while (c_loop-- != 0);
853 debug("%s:%d clocks=%u ... end\n", __func__, __LINE__, clocks);
857 * rw_mgr_mem_init_load_regs() - Load instruction registers
858 * @cntr0: Counter 0 value
859 * @cntr1: Counter 1 value
860 * @cntr2: Counter 2 value
861 * @jump: Jump instruction value
863 * Load instruction registers.
865 static void rw_mgr_mem_init_load_regs(u32 cntr0, u32 cntr1, u32 cntr2, u32 jump)
867 uint32_t grpaddr = SDR_PHYGRP_RWMGRGRP_ADDRESS |
868 RW_MGR_RUN_SINGLE_GROUP_OFFSET;
871 writel(SKIP_DELAY_LOOP_VALUE_OR_ZERO(cntr0),
872 &sdr_rw_load_mgr_regs->load_cntr0);
873 writel(SKIP_DELAY_LOOP_VALUE_OR_ZERO(cntr1),
874 &sdr_rw_load_mgr_regs->load_cntr1);
875 writel(SKIP_DELAY_LOOP_VALUE_OR_ZERO(cntr2),
876 &sdr_rw_load_mgr_regs->load_cntr2);
878 /* Load jump address */
879 writel(jump, &sdr_rw_load_jump_mgr_regs->load_jump_add0);
880 writel(jump, &sdr_rw_load_jump_mgr_regs->load_jump_add1);
881 writel(jump, &sdr_rw_load_jump_mgr_regs->load_jump_add2);
883 /* Execute count instruction */
884 writel(jump, grpaddr);
888 * rw_mgr_mem_load_user() - Load user calibration values
889 * @fin1: Final instruction 1
890 * @fin2: Final instruction 2
891 * @precharge: If 1, precharge the banks at the end
893 * Load user calibration values and optionally precharge the banks.
895 static void rw_mgr_mem_load_user(const u32 fin1, const u32 fin2,
898 u32 grpaddr = SDR_PHYGRP_RWMGRGRP_ADDRESS |
899 RW_MGR_RUN_SINGLE_GROUP_OFFSET;
902 for (r = 0; r < RW_MGR_MEM_NUMBER_OF_RANKS; r++) {
903 if (param->skip_ranks[r]) {
904 /* request to skip the rank */
909 set_rank_and_odt_mask(r, RW_MGR_ODT_MODE_OFF);
911 /* precharge all banks ... */
913 writel(RW_MGR_PRECHARGE_ALL, grpaddr);
916 * USER Use Mirror-ed commands for odd ranks if address
919 if ((RW_MGR_MEM_ADDRESS_MIRRORING >> r) & 0x1) {
920 set_jump_as_return();
921 writel(RW_MGR_MRS2_MIRR, grpaddr);
922 delay_for_n_mem_clocks(4);
923 set_jump_as_return();
924 writel(RW_MGR_MRS3_MIRR, grpaddr);
925 delay_for_n_mem_clocks(4);
926 set_jump_as_return();
927 writel(RW_MGR_MRS1_MIRR, grpaddr);
928 delay_for_n_mem_clocks(4);
929 set_jump_as_return();
930 writel(fin1, grpaddr);
932 set_jump_as_return();
933 writel(RW_MGR_MRS2, grpaddr);
934 delay_for_n_mem_clocks(4);
935 set_jump_as_return();
936 writel(RW_MGR_MRS3, grpaddr);
937 delay_for_n_mem_clocks(4);
938 set_jump_as_return();
939 writel(RW_MGR_MRS1, grpaddr);
940 set_jump_as_return();
941 writel(fin2, grpaddr);
947 set_jump_as_return();
948 writel(RW_MGR_ZQCL, grpaddr);
950 /* tZQinit = tDLLK = 512 ck cycles */
951 delay_for_n_mem_clocks(512);
956 * rw_mgr_mem_initialize() - Initialize RW Manager
958 * Initialize RW Manager.
960 static void rw_mgr_mem_initialize(void)
962 debug("%s:%d\n", __func__, __LINE__);
964 /* The reset / cke part of initialization is broadcasted to all ranks */
965 writel(RW_MGR_RANK_ALL, SDR_PHYGRP_RWMGRGRP_ADDRESS |
966 RW_MGR_SET_CS_AND_ODT_MASK_OFFSET);
969 * Here's how you load register for a loop
970 * Counters are located @ 0x800
971 * Jump address are located @ 0xC00
972 * For both, registers 0 to 3 are selected using bits 3 and 2, like
973 * in 0x800, 0x804, 0x808, 0x80C and 0xC00, 0xC04, 0xC08, 0xC0C
974 * I know this ain't pretty, but Avalon bus throws away the 2 least
978 /* Start with memory RESET activated */
983 * 200us @ 266MHz (3.75 ns) ~ 54000 clock cycles
984 * If a and b are the number of iteration in 2 nested loops
985 * it takes the following number of cycles to complete the operation:
986 * number_of_cycles = ((2 + n) * a + 2) * b
987 * where n is the number of instruction in the inner loop
988 * One possible solution is n = 0 , a = 256 , b = 106 => a = FF,
991 rw_mgr_mem_init_load_regs(SEQ_TINIT_CNTR0_VAL, SEQ_TINIT_CNTR1_VAL,
993 RW_MGR_INIT_RESET_0_CKE_0);
995 /* Indicate that memory is stable. */
996 writel(1, &phy_mgr_cfg->reset_mem_stbl);
999 * transition the RESET to high
1004 * 500us @ 266MHz (3.75 ns) ~ 134000 clock cycles
1005 * If a and b are the number of iteration in 2 nested loops
1006 * it takes the following number of cycles to complete the operation
1007 * number_of_cycles = ((2 + n) * a + 2) * b
1008 * where n is the number of instruction in the inner loop
1009 * One possible solution is n = 2 , a = 131 , b = 256 => a = 83,
1012 rw_mgr_mem_init_load_regs(SEQ_TRESET_CNTR0_VAL, SEQ_TRESET_CNTR1_VAL,
1013 SEQ_TRESET_CNTR2_VAL,
1014 RW_MGR_INIT_RESET_1_CKE_0);
1016 /* Bring up clock enable. */
1018 /* tXRP < 250 ck cycles */
1019 delay_for_n_mem_clocks(250);
1021 rw_mgr_mem_load_user(RW_MGR_MRS0_DLL_RESET_MIRR, RW_MGR_MRS0_DLL_RESET,
1026 * At the end of calibration we have to program the user settings in, and
1027 * USER hand off the memory to the user.
1029 static void rw_mgr_mem_handoff(void)
1031 rw_mgr_mem_load_user(RW_MGR_MRS0_USER_MIRR, RW_MGR_MRS0_USER, 1);
1033 * USER need to wait tMOD (12CK or 15ns) time before issuing
1034 * other commands, but we will have plenty of NIOS cycles before
1035 * actual handoff so its okay.
1040 * rw_mgr_mem_calibrate_read_test_patterns() - Read back test patterns
1041 * @rank_bgn: Rank number
1042 * @group: Read/Write Group
1043 * @all_ranks: Test all ranks
1045 * Performs a guaranteed read on the patterns we are going to use during a
1046 * read test to ensure memory works.
1049 rw_mgr_mem_calibrate_read_test_patterns(const u32 rank_bgn, const u32 group,
1050 const u32 all_ranks)
1052 const u32 addr = SDR_PHYGRP_RWMGRGRP_ADDRESS |
1053 RW_MGR_RUN_SINGLE_GROUP_OFFSET;
1054 const u32 addr_offset =
1055 (group * RW_MGR_MEM_VIRTUAL_GROUPS_PER_READ_DQS) << 2;
1056 const u32 rank_end = all_ranks ?
1057 RW_MGR_MEM_NUMBER_OF_RANKS :
1058 (rank_bgn + NUM_RANKS_PER_SHADOW_REG);
1059 const u32 shift_ratio = RW_MGR_MEM_DQ_PER_READ_DQS /
1060 RW_MGR_MEM_VIRTUAL_GROUPS_PER_READ_DQS;
1061 const u32 correct_mask_vg = param->read_correct_mask_vg;
1063 u32 tmp_bit_chk, base_rw_mgr, bit_chk;
1067 bit_chk = param->read_correct_mask;
1069 for (r = rank_bgn; r < rank_end; r++) {
1070 /* Request to skip the rank */
1071 if (param->skip_ranks[r])
1075 set_rank_and_odt_mask(r, RW_MGR_ODT_MODE_READ_WRITE);
1077 /* Load up a constant bursts of read commands */
1078 writel(0x20, &sdr_rw_load_mgr_regs->load_cntr0);
1079 writel(RW_MGR_GUARANTEED_READ,
1080 &sdr_rw_load_jump_mgr_regs->load_jump_add0);
1082 writel(0x20, &sdr_rw_load_mgr_regs->load_cntr1);
1083 writel(RW_MGR_GUARANTEED_READ_CONT,
1084 &sdr_rw_load_jump_mgr_regs->load_jump_add1);
1087 for (vg = RW_MGR_MEM_VIRTUAL_GROUPS_PER_READ_DQS - 1;
1089 /* Reset the FIFOs to get pointers to known state. */
1090 writel(0, &phy_mgr_cmd->fifo_reset);
1091 writel(0, SDR_PHYGRP_RWMGRGRP_ADDRESS |
1092 RW_MGR_RESET_READ_DATAPATH_OFFSET);
1093 writel(RW_MGR_GUARANTEED_READ,
1094 addr + addr_offset + (vg << 2));
1096 base_rw_mgr = readl(SDR_PHYGRP_RWMGRGRP_ADDRESS);
1097 tmp_bit_chk <<= shift_ratio;
1098 tmp_bit_chk |= correct_mask_vg & ~base_rw_mgr;
1101 bit_chk &= tmp_bit_chk;
1104 writel(RW_MGR_CLEAR_DQS_ENABLE, addr + (group << 2));
1106 set_rank_and_odt_mask(0, RW_MGR_ODT_MODE_OFF);
1108 if (bit_chk != param->read_correct_mask)
1111 debug_cond(DLEVEL == 1,
1112 "%s:%d test_load_patterns(%u,ALL) => (%u == %u) => %i\n",
1113 __func__, __LINE__, group, bit_chk,
1114 param->read_correct_mask, ret);
1120 * rw_mgr_mem_calibrate_read_load_patterns() - Load up the patterns for read test
1121 * @rank_bgn: Rank number
1122 * @all_ranks: Test all ranks
1124 * Load up the patterns we are going to use during a read test.
1126 static void rw_mgr_mem_calibrate_read_load_patterns(const u32 rank_bgn,
1127 const int all_ranks)
1129 const u32 rank_end = all_ranks ?
1130 RW_MGR_MEM_NUMBER_OF_RANKS :
1131 (rank_bgn + NUM_RANKS_PER_SHADOW_REG);
1134 debug("%s:%d\n", __func__, __LINE__);
1136 for (r = rank_bgn; r < rank_end; r++) {
1137 if (param->skip_ranks[r])
1138 /* request to skip the rank */
1142 set_rank_and_odt_mask(r, RW_MGR_ODT_MODE_READ_WRITE);
1144 /* Load up a constant bursts */
1145 writel(0x20, &sdr_rw_load_mgr_regs->load_cntr0);
1147 writel(RW_MGR_GUARANTEED_WRITE_WAIT0,
1148 &sdr_rw_load_jump_mgr_regs->load_jump_add0);
1150 writel(0x20, &sdr_rw_load_mgr_regs->load_cntr1);
1152 writel(RW_MGR_GUARANTEED_WRITE_WAIT1,
1153 &sdr_rw_load_jump_mgr_regs->load_jump_add1);
1155 writel(0x04, &sdr_rw_load_mgr_regs->load_cntr2);
1157 writel(RW_MGR_GUARANTEED_WRITE_WAIT2,
1158 &sdr_rw_load_jump_mgr_regs->load_jump_add2);
1160 writel(0x04, &sdr_rw_load_mgr_regs->load_cntr3);
1162 writel(RW_MGR_GUARANTEED_WRITE_WAIT3,
1163 &sdr_rw_load_jump_mgr_regs->load_jump_add3);
1165 writel(RW_MGR_GUARANTEED_WRITE, SDR_PHYGRP_RWMGRGRP_ADDRESS |
1166 RW_MGR_RUN_SINGLE_GROUP_OFFSET);
1169 set_rank_and_odt_mask(0, RW_MGR_ODT_MODE_OFF);
1173 * try a read and see if it returns correct data back. has dummy reads
1174 * inserted into the mix used to align dqs enable. has more thorough checks
1175 * than the regular read test.
1178 rw_mgr_mem_calibrate_read_test(const u32 rank_bgn, const u32 group,
1179 const u32 num_tries, const u32 all_correct,
1181 const u32 all_groups, const u32 all_ranks)
1183 const u32 rank_end = all_ranks ? RW_MGR_MEM_NUMBER_OF_RANKS :
1184 (rank_bgn + NUM_RANKS_PER_SHADOW_REG);
1185 const u32 quick_read_mode =
1186 ((STATIC_CALIB_STEPS & CALIB_SKIP_DELAY_SWEEPS) &&
1187 ENABLE_SUPER_QUICK_CALIBRATION);
1188 u32 correct_mask_vg = param->read_correct_mask_vg;
1195 *bit_chk = param->read_correct_mask;
1197 for (r = rank_bgn; r < rank_end; r++) {
1198 if (param->skip_ranks[r])
1199 /* request to skip the rank */
1203 set_rank_and_odt_mask(r, RW_MGR_ODT_MODE_READ_WRITE);
1205 writel(0x10, &sdr_rw_load_mgr_regs->load_cntr1);
1207 writel(RW_MGR_READ_B2B_WAIT1,
1208 &sdr_rw_load_jump_mgr_regs->load_jump_add1);
1210 writel(0x10, &sdr_rw_load_mgr_regs->load_cntr2);
1211 writel(RW_MGR_READ_B2B_WAIT2,
1212 &sdr_rw_load_jump_mgr_regs->load_jump_add2);
1214 if (quick_read_mode)
1215 writel(0x1, &sdr_rw_load_mgr_regs->load_cntr0);
1216 /* need at least two (1+1) reads to capture failures */
1217 else if (all_groups)
1218 writel(0x06, &sdr_rw_load_mgr_regs->load_cntr0);
1220 writel(0x32, &sdr_rw_load_mgr_regs->load_cntr0);
1222 writel(RW_MGR_READ_B2B,
1223 &sdr_rw_load_jump_mgr_regs->load_jump_add0);
1225 writel(RW_MGR_MEM_IF_READ_DQS_WIDTH *
1226 RW_MGR_MEM_VIRTUAL_GROUPS_PER_READ_DQS - 1,
1227 &sdr_rw_load_mgr_regs->load_cntr3);
1229 writel(0x0, &sdr_rw_load_mgr_regs->load_cntr3);
1231 writel(RW_MGR_READ_B2B,
1232 &sdr_rw_load_jump_mgr_regs->load_jump_add3);
1235 for (vg = RW_MGR_MEM_VIRTUAL_GROUPS_PER_READ_DQS - 1; vg >= 0;
1237 /* reset the fifos to get pointers to known state */
1238 writel(0, &phy_mgr_cmd->fifo_reset);
1239 writel(0, SDR_PHYGRP_RWMGRGRP_ADDRESS |
1240 RW_MGR_RESET_READ_DATAPATH_OFFSET);
1242 tmp_bit_chk = tmp_bit_chk << (RW_MGR_MEM_DQ_PER_READ_DQS
1243 / RW_MGR_MEM_VIRTUAL_GROUPS_PER_READ_DQS);
1246 addr = SDR_PHYGRP_RWMGRGRP_ADDRESS | RW_MGR_RUN_ALL_GROUPS_OFFSET;
1248 addr = SDR_PHYGRP_RWMGRGRP_ADDRESS | RW_MGR_RUN_SINGLE_GROUP_OFFSET;
1250 writel(RW_MGR_READ_B2B, addr +
1251 ((group * RW_MGR_MEM_VIRTUAL_GROUPS_PER_READ_DQS +
1254 base_rw_mgr = readl(SDR_PHYGRP_RWMGRGRP_ADDRESS);
1255 tmp_bit_chk = tmp_bit_chk | (correct_mask_vg & ~(base_rw_mgr));
1258 *bit_chk &= tmp_bit_chk;
1261 addr = SDR_PHYGRP_RWMGRGRP_ADDRESS | RW_MGR_RUN_SINGLE_GROUP_OFFSET;
1262 writel(RW_MGR_CLEAR_DQS_ENABLE, addr + (group << 2));
1264 set_rank_and_odt_mask(0, RW_MGR_ODT_MODE_OFF);
1267 ret = (*bit_chk == param->read_correct_mask);
1268 debug_cond(DLEVEL == 2,
1269 "%s:%d read_test(%u,ALL,%u) => (%u == %u) => %i\n",
1270 __func__, __LINE__, group, all_groups, *bit_chk,
1271 param->read_correct_mask, ret);
1273 ret = (*bit_chk != 0x00);
1274 debug_cond(DLEVEL == 2,
1275 "%s:%d read_test(%u,ONE,%u) => (%u != %u) => %i\n",
1276 __func__, __LINE__, group, all_groups, *bit_chk,
1284 * rw_mgr_mem_calibrate_read_test_all_ranks() - Perform READ test on all ranks
1285 * @grp: Read/Write group
1286 * @num_tries: Number of retries of the test
1287 * @all_correct: All bits must be correct in the mask
1288 * @all_groups: Test all R/W groups
1290 * Perform a READ test across all memory ranks.
1293 rw_mgr_mem_calibrate_read_test_all_ranks(const u32 grp, const u32 num_tries,
1294 const u32 all_correct,
1295 const u32 all_groups)
1298 return rw_mgr_mem_calibrate_read_test(0, grp, num_tries, all_correct,
1299 &bit_chk, all_groups, 1);
1303 * rw_mgr_incr_vfifo() - Increase VFIFO value
1304 * @grp: Read/Write group
1306 * Increase VFIFO value.
1308 static void rw_mgr_incr_vfifo(const u32 grp)
1310 writel(grp, &phy_mgr_cmd->inc_vfifo_hard_phy);
1314 * rw_mgr_decr_vfifo() - Decrease VFIFO value
1315 * @grp: Read/Write group
1317 * Decrease VFIFO value.
1319 static void rw_mgr_decr_vfifo(const u32 grp)
1323 for (i = 0; i < VFIFO_SIZE - 1; i++)
1324 rw_mgr_incr_vfifo(grp);
1328 * find_vfifo_failing_read() - Push VFIFO to get a failing read
1329 * @grp: Read/Write group
1331 * Push VFIFO until a failing read happens.
1333 static int find_vfifo_failing_read(const u32 grp)
1335 u32 v, ret, fail_cnt = 0;
1337 for (v = 0; v < VFIFO_SIZE; v++) {
1338 debug_cond(DLEVEL == 2, "%s:%d: vfifo %u\n",
1339 __func__, __LINE__, v);
1340 ret = rw_mgr_mem_calibrate_read_test_all_ranks(grp, 1,
1349 /* Fiddle with FIFO. */
1350 rw_mgr_incr_vfifo(grp);
1353 /* No failing read found! Something must have gone wrong. */
1354 debug_cond(DLEVEL == 2, "%s:%d: vfifo failed\n", __func__, __LINE__);
1359 * sdr_find_phase_delay() - Find DQS enable phase or delay
1360 * @working: If 1, look for working phase/delay, if 0, look for non-working
1361 * @delay: If 1, look for delay, if 0, look for phase
1362 * @grp: Read/Write group
1363 * @work: Working window position
1364 * @work_inc: Working window increment
1365 * @pd: DQS Phase/Delay Iterator
1367 * Find working or non-working DQS enable phase setting.
1369 static int sdr_find_phase_delay(int working, int delay, const u32 grp,
1370 u32 *work, const u32 work_inc, u32 *pd)
1372 const u32 max = delay ? IO_DQS_EN_DELAY_MAX : IO_DQS_EN_PHASE_MAX;
1375 for (; *pd <= max; (*pd)++) {
1377 scc_mgr_set_dqs_en_delay_all_ranks(grp, *pd);
1379 scc_mgr_set_dqs_en_phase_all_ranks(grp, *pd);
1381 ret = rw_mgr_mem_calibrate_read_test_all_ranks(grp, 1,
1396 * sdr_find_phase() - Find DQS enable phase
1397 * @working: If 1, look for working phase, if 0, look for non-working phase
1398 * @grp: Read/Write group
1399 * @work: Working window position
1401 * @p: DQS Phase Iterator
1403 * Find working or non-working DQS enable phase setting.
1405 static int sdr_find_phase(int working, const u32 grp, u32 *work,
1408 const u32 end = VFIFO_SIZE + (working ? 0 : 1);
1411 for (; *i < end; (*i)++) {
1415 ret = sdr_find_phase_delay(working, 0, grp, work,
1416 IO_DELAY_PER_OPA_TAP, p);
1420 if (*p > IO_DQS_EN_PHASE_MAX) {
1421 /* Fiddle with FIFO. */
1422 rw_mgr_incr_vfifo(grp);
1432 * sdr_working_phase() - Find working DQS enable phase
1433 * @grp: Read/Write group
1434 * @work_bgn: Working window start position
1435 * @d: dtaps output value
1436 * @p: DQS Phase Iterator
1439 * Find working DQS enable phase setting.
1441 static int sdr_working_phase(const u32 grp, u32 *work_bgn, u32 *d,
1444 const u32 dtaps_per_ptap = IO_DELAY_PER_OPA_TAP /
1445 IO_DELAY_PER_DQS_EN_DCHAIN_TAP;
1450 for (*d = 0; *d <= dtaps_per_ptap; (*d)++) {
1452 scc_mgr_set_dqs_en_delay_all_ranks(grp, *d);
1453 ret = sdr_find_phase(1, grp, work_bgn, i, p);
1456 *work_bgn += IO_DELAY_PER_DQS_EN_DCHAIN_TAP;
1459 /* Cannot find working solution */
1460 debug_cond(DLEVEL == 2, "%s:%d find_dqs_en_phase: no vfifo/ptap/dtap\n",
1461 __func__, __LINE__);
1466 * sdr_backup_phase() - Find DQS enable backup phase
1467 * @grp: Read/Write group
1468 * @work_bgn: Working window start position
1469 * @p: DQS Phase Iterator
1471 * Find DQS enable backup phase setting.
1473 static void sdr_backup_phase(const u32 grp, u32 *work_bgn, u32 *p)
1478 /* Special case code for backing up a phase */
1480 *p = IO_DQS_EN_PHASE_MAX;
1481 rw_mgr_decr_vfifo(grp);
1485 tmp_delay = *work_bgn - IO_DELAY_PER_OPA_TAP;
1486 scc_mgr_set_dqs_en_phase_all_ranks(grp, *p);
1488 for (d = 0; d <= IO_DQS_EN_DELAY_MAX && tmp_delay < *work_bgn; d++) {
1489 scc_mgr_set_dqs_en_delay_all_ranks(grp, d);
1491 ret = rw_mgr_mem_calibrate_read_test_all_ranks(grp, 1,
1494 *work_bgn = tmp_delay;
1498 tmp_delay += IO_DELAY_PER_DQS_EN_DCHAIN_TAP;
1501 /* Restore VFIFO to old state before we decremented it (if needed). */
1503 if (*p > IO_DQS_EN_PHASE_MAX) {
1505 rw_mgr_incr_vfifo(grp);
1508 scc_mgr_set_dqs_en_delay_all_ranks(grp, 0);
1512 * sdr_nonworking_phase() - Find non-working DQS enable phase
1513 * @grp: Read/Write group
1514 * @work_end: Working window end position
1515 * @p: DQS Phase Iterator
1518 * Find non-working DQS enable phase setting.
1520 static int sdr_nonworking_phase(const u32 grp, u32 *work_end, u32 *p, u32 *i)
1525 *work_end += IO_DELAY_PER_OPA_TAP;
1526 if (*p > IO_DQS_EN_PHASE_MAX) {
1527 /* Fiddle with FIFO. */
1529 rw_mgr_incr_vfifo(grp);
1532 ret = sdr_find_phase(0, grp, work_end, i, p);
1534 /* Cannot see edge of failing read. */
1535 debug_cond(DLEVEL == 2, "%s:%d: end: failed\n",
1536 __func__, __LINE__);
1543 * sdr_find_window_center() - Find center of the working DQS window.
1544 * @grp: Read/Write group
1545 * @work_bgn: First working settings
1546 * @work_end: Last working settings
1548 * Find center of the working DQS enable window.
1550 static int sdr_find_window_center(const u32 grp, const u32 work_bgn,
1557 work_mid = (work_bgn + work_end) / 2;
1559 debug_cond(DLEVEL == 2, "work_bgn=%d work_end=%d work_mid=%d\n",
1560 work_bgn, work_end, work_mid);
1561 /* Get the middle delay to be less than a VFIFO delay */
1562 tmp_delay = (IO_DQS_EN_PHASE_MAX + 1) * IO_DELAY_PER_OPA_TAP;
1564 debug_cond(DLEVEL == 2, "vfifo ptap delay %d\n", tmp_delay);
1565 work_mid %= tmp_delay;
1566 debug_cond(DLEVEL == 2, "new work_mid %d\n", work_mid);
1568 tmp_delay = rounddown(work_mid, IO_DELAY_PER_OPA_TAP);
1569 if (tmp_delay > IO_DQS_EN_PHASE_MAX * IO_DELAY_PER_OPA_TAP)
1570 tmp_delay = IO_DQS_EN_PHASE_MAX * IO_DELAY_PER_OPA_TAP;
1571 p = tmp_delay / IO_DELAY_PER_OPA_TAP;
1573 debug_cond(DLEVEL == 2, "new p %d, tmp_delay=%d\n", p, tmp_delay);
1575 d = DIV_ROUND_UP(work_mid - tmp_delay, IO_DELAY_PER_DQS_EN_DCHAIN_TAP);
1576 if (d > IO_DQS_EN_DELAY_MAX)
1577 d = IO_DQS_EN_DELAY_MAX;
1578 tmp_delay += d * IO_DELAY_PER_DQS_EN_DCHAIN_TAP;
1580 debug_cond(DLEVEL == 2, "new d %d, tmp_delay=%d\n", d, tmp_delay);
1582 scc_mgr_set_dqs_en_phase_all_ranks(grp, p);
1583 scc_mgr_set_dqs_en_delay_all_ranks(grp, d);
1586 * push vfifo until we can successfully calibrate. We can do this
1587 * because the largest possible margin in 1 VFIFO cycle.
1589 for (i = 0; i < VFIFO_SIZE; i++) {
1590 debug_cond(DLEVEL == 2, "find_dqs_en_phase: center\n");
1591 if (rw_mgr_mem_calibrate_read_test_all_ranks(grp, 1,
1594 debug_cond(DLEVEL == 2,
1595 "%s:%d center: found: ptap=%u dtap=%u\n",
1596 __func__, __LINE__, p, d);
1600 /* Fiddle with FIFO. */
1601 rw_mgr_incr_vfifo(grp);
1604 debug_cond(DLEVEL == 2, "%s:%d center: failed.\n",
1605 __func__, __LINE__);
1610 * rw_mgr_mem_calibrate_vfifo_find_dqs_en_phase() - Find a good DQS enable to use
1611 * @grp: Read/Write Group
1613 * Find a good DQS enable to use.
1615 static int rw_mgr_mem_calibrate_vfifo_find_dqs_en_phase(const u32 grp)
1619 u32 work_bgn, work_end;
1620 u32 found_passing_read, found_failing_read, initial_failing_dtap;
1623 debug("%s:%d %u\n", __func__, __LINE__, grp);
1625 reg_file_set_sub_stage(CAL_SUBSTAGE_VFIFO_CENTER);
1627 scc_mgr_set_dqs_en_delay_all_ranks(grp, 0);
1628 scc_mgr_set_dqs_en_phase_all_ranks(grp, 0);
1630 /* Step 0: Determine number of delay taps for each phase tap. */
1631 dtaps_per_ptap = IO_DELAY_PER_OPA_TAP / IO_DELAY_PER_DQS_EN_DCHAIN_TAP;
1633 /* Step 1: First push vfifo until we get a failing read. */
1634 find_vfifo_failing_read(grp);
1636 /* Step 2: Find first working phase, increment in ptaps. */
1638 ret = sdr_working_phase(grp, &work_bgn, &d, &p, &i);
1642 work_end = work_bgn;
1645 * If d is 0 then the working window covers a phase tap and we can
1646 * follow the old procedure. Otherwise, we've found the beginning
1647 * and we need to increment the dtaps until we find the end.
1651 * Step 3a: If we have room, back off by one and
1652 * increment in dtaps.
1654 sdr_backup_phase(grp, &work_bgn, &p);
1657 * Step 4a: go forward from working phase to non working
1658 * phase, increment in ptaps.
1660 ret = sdr_nonworking_phase(grp, &work_end, &p, &i);
1664 /* Step 5a: Back off one from last, increment in dtaps. */
1666 /* Special case code for backing up a phase */
1668 p = IO_DQS_EN_PHASE_MAX;
1669 rw_mgr_decr_vfifo(grp);
1674 work_end -= IO_DELAY_PER_OPA_TAP;
1675 scc_mgr_set_dqs_en_phase_all_ranks(grp, p);
1679 debug_cond(DLEVEL == 2, "%s:%d p: ptap=%u\n",
1680 __func__, __LINE__, p);
1683 /* The dtap increment to find the failing edge is done here. */
1684 sdr_find_phase_delay(0, 1, grp, &work_end,
1685 IO_DELAY_PER_DQS_EN_DCHAIN_TAP, &d);
1687 /* Go back to working dtap */
1689 work_end -= IO_DELAY_PER_DQS_EN_DCHAIN_TAP;
1691 debug_cond(DLEVEL == 2,
1692 "%s:%d p/d: ptap=%u dtap=%u end=%u\n",
1693 __func__, __LINE__, p, d - 1, work_end);
1695 if (work_end < work_bgn) {
1697 debug_cond(DLEVEL == 2, "%s:%d end-2: failed\n",
1698 __func__, __LINE__);
1702 debug_cond(DLEVEL == 2, "%s:%d found range [%u,%u]\n",
1703 __func__, __LINE__, work_bgn, work_end);
1706 * We need to calculate the number of dtaps that equal a ptap.
1707 * To do that we'll back up a ptap and re-find the edge of the
1708 * window using dtaps
1710 debug_cond(DLEVEL == 2, "%s:%d calculate dtaps_per_ptap for tracking\n",
1711 __func__, __LINE__);
1713 /* Special case code for backing up a phase */
1715 p = IO_DQS_EN_PHASE_MAX;
1716 rw_mgr_decr_vfifo(grp);
1717 debug_cond(DLEVEL == 2, "%s:%d backedup cycle/phase: p=%u\n",
1718 __func__, __LINE__, p);
1721 debug_cond(DLEVEL == 2, "%s:%d backedup phase only: p=%u",
1722 __func__, __LINE__, p);
1725 scc_mgr_set_dqs_en_phase_all_ranks(grp, p);
1728 * Increase dtap until we first see a passing read (in case the
1729 * window is smaller than a ptap), and then a failing read to
1730 * mark the edge of the window again.
1733 /* Find a passing read. */
1734 debug_cond(DLEVEL == 2, "%s:%d find passing read\n",
1735 __func__, __LINE__);
1737 initial_failing_dtap = d;
1739 found_passing_read = !sdr_find_phase_delay(1, 1, grp, NULL, 0, &d);
1740 if (found_passing_read) {
1741 /* Find a failing read. */
1742 debug_cond(DLEVEL == 2, "%s:%d find failing read\n",
1743 __func__, __LINE__);
1745 found_failing_read = !sdr_find_phase_delay(0, 1, grp, NULL, 0,
1748 debug_cond(DLEVEL == 1,
1749 "%s:%d failed to calculate dtaps per ptap. Fall back on static value\n",
1750 __func__, __LINE__);
1754 * The dynamically calculated dtaps_per_ptap is only valid if we
1755 * found a passing/failing read. If we didn't, it means d hit the max
1756 * (IO_DQS_EN_DELAY_MAX). Otherwise, dtaps_per_ptap retains its
1757 * statically calculated value.
1759 if (found_passing_read && found_failing_read)
1760 dtaps_per_ptap = d - initial_failing_dtap;
1762 writel(dtaps_per_ptap, &sdr_reg_file->dtaps_per_ptap);
1763 debug_cond(DLEVEL == 2, "%s:%d dtaps_per_ptap=%u - %u = %u",
1764 __func__, __LINE__, d, initial_failing_dtap, dtaps_per_ptap);
1766 /* Step 6: Find the centre of the window. */
1767 ret = sdr_find_window_center(grp, work_bgn, work_end);
1772 /* per-bit deskew DQ and center */
1773 static uint32_t rw_mgr_mem_calibrate_vfifo_center(uint32_t rank_bgn,
1774 uint32_t write_group, uint32_t read_group, uint32_t test_bgn,
1775 uint32_t use_read_test, uint32_t update_fom)
1777 uint32_t i, p, d, min_index;
1779 * Store these as signed since there are comparisons with
1783 uint32_t sticky_bit_chk;
1784 int32_t left_edge[RW_MGR_MEM_DQ_PER_READ_DQS];
1785 int32_t right_edge[RW_MGR_MEM_DQ_PER_READ_DQS];
1786 int32_t final_dq[RW_MGR_MEM_DQ_PER_READ_DQS];
1788 int32_t orig_mid_min, mid_min;
1789 int32_t new_dqs, start_dqs, start_dqs_en, shift_dq, final_dqs,
1791 int32_t dq_margin, dqs_margin;
1793 uint32_t temp_dq_in_delay1, temp_dq_in_delay2;
1796 debug("%s:%d: %u %u", __func__, __LINE__, read_group, test_bgn);
1798 addr = SDR_PHYGRP_SCCGRP_ADDRESS | SCC_MGR_DQS_IN_DELAY_OFFSET;
1799 start_dqs = readl(addr + (read_group << 2));
1800 if (IO_SHIFT_DQS_EN_WHEN_SHIFT_DQS)
1801 start_dqs_en = readl(addr + ((read_group << 2)
1802 - IO_DQS_EN_DELAY_OFFSET));
1804 /* set the left and right edge of each bit to an illegal value */
1805 /* use (IO_IO_IN_DELAY_MAX + 1) as an illegal value */
1807 for (i = 0; i < RW_MGR_MEM_DQ_PER_READ_DQS; i++) {
1808 left_edge[i] = IO_IO_IN_DELAY_MAX + 1;
1809 right_edge[i] = IO_IO_IN_DELAY_MAX + 1;
1812 /* Search for the left edge of the window for each bit */
1813 for (d = 0; d <= IO_IO_IN_DELAY_MAX; d++) {
1814 scc_mgr_apply_group_dq_in_delay(write_group, test_bgn, d);
1816 writel(0, &sdr_scc_mgr->update);
1819 * Stop searching when the read test doesn't pass AND when
1820 * we've seen a passing read on every bit.
1822 if (use_read_test) {
1823 stop = !rw_mgr_mem_calibrate_read_test(rank_bgn,
1824 read_group, NUM_READ_PB_TESTS, PASS_ONE_BIT,
1827 rw_mgr_mem_calibrate_write_test(rank_bgn, write_group,
1830 bit_chk = bit_chk >> (RW_MGR_MEM_DQ_PER_READ_DQS *
1831 (read_group - (write_group *
1832 RW_MGR_MEM_IF_READ_DQS_WIDTH /
1833 RW_MGR_MEM_IF_WRITE_DQS_WIDTH)));
1834 stop = (bit_chk == 0);
1836 sticky_bit_chk = sticky_bit_chk | bit_chk;
1837 stop = stop && (sticky_bit_chk == param->read_correct_mask);
1838 debug_cond(DLEVEL == 2, "%s:%d vfifo_center(left): dtap=%u => %u == %u \
1839 && %u", __func__, __LINE__, d,
1841 param->read_correct_mask, stop);
1846 for (i = 0; i < RW_MGR_MEM_DQ_PER_READ_DQS; i++) {
1848 /* Remember a passing test as the
1852 /* If a left edge has not been seen yet,
1853 then a future passing test will mark
1854 this edge as the right edge */
1856 IO_IO_IN_DELAY_MAX + 1) {
1857 right_edge[i] = -(d + 1);
1860 bit_chk = bit_chk >> 1;
1865 /* Reset DQ delay chains to 0 */
1866 scc_mgr_apply_group_dq_in_delay(test_bgn, 0);
1868 for (i = RW_MGR_MEM_DQ_PER_READ_DQS - 1;; i--) {
1869 debug_cond(DLEVEL == 2, "%s:%d vfifo_center: left_edge[%u]: \
1870 %d right_edge[%u]: %d\n", __func__, __LINE__,
1871 i, left_edge[i], i, right_edge[i]);
1874 * Check for cases where we haven't found the left edge,
1875 * which makes our assignment of the the right edge invalid.
1876 * Reset it to the illegal value.
1878 if ((left_edge[i] == IO_IO_IN_DELAY_MAX + 1) && (
1879 right_edge[i] != IO_IO_IN_DELAY_MAX + 1)) {
1880 right_edge[i] = IO_IO_IN_DELAY_MAX + 1;
1881 debug_cond(DLEVEL == 2, "%s:%d vfifo_center: reset \
1882 right_edge[%u]: %d\n", __func__, __LINE__,
1887 * Reset sticky bit (except for bits where we have seen
1888 * both the left and right edge).
1890 sticky_bit_chk = sticky_bit_chk << 1;
1891 if ((left_edge[i] != IO_IO_IN_DELAY_MAX + 1) &&
1892 (right_edge[i] != IO_IO_IN_DELAY_MAX + 1)) {
1893 sticky_bit_chk = sticky_bit_chk | 1;
1900 /* Search for the right edge of the window for each bit */
1901 for (d = 0; d <= IO_DQS_IN_DELAY_MAX - start_dqs; d++) {
1902 scc_mgr_set_dqs_bus_in_delay(read_group, d + start_dqs);
1903 if (IO_SHIFT_DQS_EN_WHEN_SHIFT_DQS) {
1904 uint32_t delay = d + start_dqs_en;
1905 if (delay > IO_DQS_EN_DELAY_MAX)
1906 delay = IO_DQS_EN_DELAY_MAX;
1907 scc_mgr_set_dqs_en_delay(read_group, delay);
1909 scc_mgr_load_dqs(read_group);
1911 writel(0, &sdr_scc_mgr->update);
1914 * Stop searching when the read test doesn't pass AND when
1915 * we've seen a passing read on every bit.
1917 if (use_read_test) {
1918 stop = !rw_mgr_mem_calibrate_read_test(rank_bgn,
1919 read_group, NUM_READ_PB_TESTS, PASS_ONE_BIT,
1922 rw_mgr_mem_calibrate_write_test(rank_bgn, write_group,
1925 bit_chk = bit_chk >> (RW_MGR_MEM_DQ_PER_READ_DQS *
1926 (read_group - (write_group *
1927 RW_MGR_MEM_IF_READ_DQS_WIDTH /
1928 RW_MGR_MEM_IF_WRITE_DQS_WIDTH)));
1929 stop = (bit_chk == 0);
1931 sticky_bit_chk = sticky_bit_chk | bit_chk;
1932 stop = stop && (sticky_bit_chk == param->read_correct_mask);
1934 debug_cond(DLEVEL == 2, "%s:%d vfifo_center(right): dtap=%u => %u == \
1935 %u && %u", __func__, __LINE__, d,
1936 sticky_bit_chk, param->read_correct_mask, stop);
1941 for (i = 0; i < RW_MGR_MEM_DQ_PER_READ_DQS; i++) {
1943 /* Remember a passing test as
1948 /* If a right edge has not been
1949 seen yet, then a future passing
1950 test will mark this edge as the
1952 if (right_edge[i] ==
1953 IO_IO_IN_DELAY_MAX + 1) {
1954 left_edge[i] = -(d + 1);
1957 /* d = 0 failed, but it passed
1958 when testing the left edge,
1959 so it must be marginal,
1961 if (right_edge[i] ==
1962 IO_IO_IN_DELAY_MAX + 1 &&
1968 /* If a right edge has not been
1969 seen yet, then a future passing
1970 test will mark this edge as the
1972 else if (right_edge[i] ==
1973 IO_IO_IN_DELAY_MAX +
1975 left_edge[i] = -(d + 1);
1980 debug_cond(DLEVEL == 2, "%s:%d vfifo_center[r,\
1981 d=%u]: ", __func__, __LINE__, d);
1982 debug_cond(DLEVEL == 2, "bit_chk_test=%d left_edge[%u]: %d ",
1983 (int)(bit_chk & 1), i, left_edge[i]);
1984 debug_cond(DLEVEL == 2, "right_edge[%u]: %d\n", i,
1986 bit_chk = bit_chk >> 1;
1991 /* Check that all bits have a window */
1992 for (i = 0; i < RW_MGR_MEM_DQ_PER_READ_DQS; i++) {
1993 debug_cond(DLEVEL == 2, "%s:%d vfifo_center: left_edge[%u]: \
1994 %d right_edge[%u]: %d", __func__, __LINE__,
1995 i, left_edge[i], i, right_edge[i]);
1996 if ((left_edge[i] == IO_IO_IN_DELAY_MAX + 1) || (right_edge[i]
1997 == IO_IO_IN_DELAY_MAX + 1)) {
1999 * Restore delay chain settings before letting the loop
2000 * in rw_mgr_mem_calibrate_vfifo to retry different
2001 * dqs/ck relationships.
2003 scc_mgr_set_dqs_bus_in_delay(read_group, start_dqs);
2004 if (IO_SHIFT_DQS_EN_WHEN_SHIFT_DQS) {
2005 scc_mgr_set_dqs_en_delay(read_group,
2008 scc_mgr_load_dqs(read_group);
2009 writel(0, &sdr_scc_mgr->update);
2011 debug_cond(DLEVEL == 1, "%s:%d vfifo_center: failed to \
2012 find edge [%u]: %d %d", __func__, __LINE__,
2013 i, left_edge[i], right_edge[i]);
2014 if (use_read_test) {
2015 set_failing_group_stage(read_group *
2016 RW_MGR_MEM_DQ_PER_READ_DQS + i,
2018 CAL_SUBSTAGE_VFIFO_CENTER);
2020 set_failing_group_stage(read_group *
2021 RW_MGR_MEM_DQ_PER_READ_DQS + i,
2022 CAL_STAGE_VFIFO_AFTER_WRITES,
2023 CAL_SUBSTAGE_VFIFO_CENTER);
2029 /* Find middle of window for each DQ bit */
2030 mid_min = left_edge[0] - right_edge[0];
2032 for (i = 1; i < RW_MGR_MEM_DQ_PER_READ_DQS; i++) {
2033 mid = left_edge[i] - right_edge[i];
2034 if (mid < mid_min) {
2041 * -mid_min/2 represents the amount that we need to move DQS.
2042 * If mid_min is odd and positive we'll need to add one to
2043 * make sure the rounding in further calculations is correct
2044 * (always bias to the right), so just add 1 for all positive values.
2049 mid_min = mid_min / 2;
2051 debug_cond(DLEVEL == 1, "%s:%d vfifo_center: mid_min=%d (index=%u)\n",
2052 __func__, __LINE__, mid_min, min_index);
2054 /* Determine the amount we can change DQS (which is -mid_min) */
2055 orig_mid_min = mid_min;
2056 new_dqs = start_dqs - mid_min;
2057 if (new_dqs > IO_DQS_IN_DELAY_MAX)
2058 new_dqs = IO_DQS_IN_DELAY_MAX;
2059 else if (new_dqs < 0)
2062 mid_min = start_dqs - new_dqs;
2063 debug_cond(DLEVEL == 1, "vfifo_center: new mid_min=%d new_dqs=%d\n",
2066 if (IO_SHIFT_DQS_EN_WHEN_SHIFT_DQS) {
2067 if (start_dqs_en - mid_min > IO_DQS_EN_DELAY_MAX)
2068 mid_min += start_dqs_en - mid_min - IO_DQS_EN_DELAY_MAX;
2069 else if (start_dqs_en - mid_min < 0)
2070 mid_min += start_dqs_en - mid_min;
2072 new_dqs = start_dqs - mid_min;
2074 debug_cond(DLEVEL == 1, "vfifo_center: start_dqs=%d start_dqs_en=%d \
2075 new_dqs=%d mid_min=%d\n", start_dqs,
2076 IO_SHIFT_DQS_EN_WHEN_SHIFT_DQS ? start_dqs_en : -1,
2079 /* Initialize data for export structures */
2080 dqs_margin = IO_IO_IN_DELAY_MAX + 1;
2081 dq_margin = IO_IO_IN_DELAY_MAX + 1;
2083 /* add delay to bring centre of all DQ windows to the same "level" */
2084 for (i = 0, p = test_bgn; i < RW_MGR_MEM_DQ_PER_READ_DQS; i++, p++) {
2085 /* Use values before divide by 2 to reduce round off error */
2086 shift_dq = (left_edge[i] - right_edge[i] -
2087 (left_edge[min_index] - right_edge[min_index]))/2 +
2088 (orig_mid_min - mid_min);
2090 debug_cond(DLEVEL == 2, "vfifo_center: before: \
2091 shift_dq[%u]=%d\n", i, shift_dq);
2093 addr = SDR_PHYGRP_SCCGRP_ADDRESS | SCC_MGR_IO_IN_DELAY_OFFSET;
2094 temp_dq_in_delay1 = readl(addr + (p << 2));
2095 temp_dq_in_delay2 = readl(addr + (i << 2));
2097 if (shift_dq + (int32_t)temp_dq_in_delay1 >
2098 (int32_t)IO_IO_IN_DELAY_MAX) {
2099 shift_dq = (int32_t)IO_IO_IN_DELAY_MAX - temp_dq_in_delay2;
2100 } else if (shift_dq + (int32_t)temp_dq_in_delay1 < 0) {
2101 shift_dq = -(int32_t)temp_dq_in_delay1;
2103 debug_cond(DLEVEL == 2, "vfifo_center: after: \
2104 shift_dq[%u]=%d\n", i, shift_dq);
2105 final_dq[i] = temp_dq_in_delay1 + shift_dq;
2106 scc_mgr_set_dq_in_delay(p, final_dq[i]);
2109 debug_cond(DLEVEL == 2, "vfifo_center: margin[%u]=[%d,%d]\n", i,
2110 left_edge[i] - shift_dq + (-mid_min),
2111 right_edge[i] + shift_dq - (-mid_min));
2112 /* To determine values for export structures */
2113 if (left_edge[i] - shift_dq + (-mid_min) < dq_margin)
2114 dq_margin = left_edge[i] - shift_dq + (-mid_min);
2116 if (right_edge[i] + shift_dq - (-mid_min) < dqs_margin)
2117 dqs_margin = right_edge[i] + shift_dq - (-mid_min);
2120 final_dqs = new_dqs;
2121 if (IO_SHIFT_DQS_EN_WHEN_SHIFT_DQS)
2122 final_dqs_en = start_dqs_en - mid_min;
2125 if (IO_SHIFT_DQS_EN_WHEN_SHIFT_DQS) {
2126 scc_mgr_set_dqs_en_delay(read_group, final_dqs_en);
2127 scc_mgr_load_dqs(read_group);
2131 scc_mgr_set_dqs_bus_in_delay(read_group, final_dqs);
2132 scc_mgr_load_dqs(read_group);
2133 debug_cond(DLEVEL == 2, "%s:%d vfifo_center: dq_margin=%d \
2134 dqs_margin=%d", __func__, __LINE__,
2135 dq_margin, dqs_margin);
2138 * Do not remove this line as it makes sure all of our decisions
2139 * have been applied. Apply the update bit.
2141 writel(0, &sdr_scc_mgr->update);
2143 return (dq_margin >= 0) && (dqs_margin >= 0);
2147 * rw_mgr_mem_calibrate_guaranteed_write() - Perform guaranteed write into the device
2148 * @rw_group: Read/Write Group
2149 * @phase: DQ/DQS phase
2151 * Because initially no communication ca be reliably performed with the memory
2152 * device, the sequencer uses a guaranteed write mechanism to write data into
2153 * the memory device.
2155 static int rw_mgr_mem_calibrate_guaranteed_write(const u32 rw_group,
2160 /* Set a particular DQ/DQS phase. */
2161 scc_mgr_set_dqdqs_output_phase_all_ranks(rw_group, phase);
2163 debug_cond(DLEVEL == 1, "%s:%d guaranteed write: g=%u p=%u\n",
2164 __func__, __LINE__, rw_group, phase);
2167 * Altera EMI_RM 2015.05.04 :: Figure 1-25
2168 * Load up the patterns used by read calibration using the
2169 * current DQDQS phase.
2171 rw_mgr_mem_calibrate_read_load_patterns(0, 1);
2173 if (gbl->phy_debug_mode_flags & PHY_DEBUG_DISABLE_GUARANTEED_READ)
2177 * Altera EMI_RM 2015.05.04 :: Figure 1-26
2178 * Back-to-Back reads of the patterns used for calibration.
2180 ret = rw_mgr_mem_calibrate_read_test_patterns(0, rw_group, 1);
2182 debug_cond(DLEVEL == 1,
2183 "%s:%d Guaranteed read test failed: g=%u p=%u\n",
2184 __func__, __LINE__, rw_group, phase);
2189 * rw_mgr_mem_calibrate_dqs_enable_calibration() - DQS Enable Calibration
2190 * @rw_group: Read/Write Group
2191 * @test_bgn: Rank at which the test begins
2193 * DQS enable calibration ensures reliable capture of the DQ signal without
2194 * glitches on the DQS line.
2196 static int rw_mgr_mem_calibrate_dqs_enable_calibration(const u32 rw_group,
2200 * Altera EMI_RM 2015.05.04 :: Figure 1-27
2201 * DQS and DQS Eanble Signal Relationships.
2204 /* We start at zero, so have one less dq to devide among */
2205 const u32 delay_step = IO_IO_IN_DELAY_MAX /
2206 (RW_MGR_MEM_DQ_PER_READ_DQS - 1);
2210 debug("%s:%d (%u,%u)\n", __func__, __LINE__, rw_group, test_bgn);
2212 /* Try different dq_in_delays since the DQ path is shorter than DQS. */
2213 for (r = 0; r < RW_MGR_MEM_NUMBER_OF_RANKS;
2214 r += NUM_RANKS_PER_SHADOW_REG) {
2215 for (i = 0, p = test_bgn, d = 0;
2216 i < RW_MGR_MEM_DQ_PER_READ_DQS;
2217 i++, p++, d += delay_step) {
2218 debug_cond(DLEVEL == 1,
2219 "%s:%d: g=%u r=%u i=%u p=%u d=%u\n",
2220 __func__, __LINE__, rw_group, r, i, p, d);
2222 scc_mgr_set_dq_in_delay(p, d);
2226 writel(0, &sdr_scc_mgr->update);
2230 * Try rw_mgr_mem_calibrate_vfifo_find_dqs_en_phase across different
2231 * dq_in_delay values
2233 ret = rw_mgr_mem_calibrate_vfifo_find_dqs_en_phase(rw_group);
2235 debug_cond(DLEVEL == 1,
2236 "%s:%d: g=%u found=%u; Reseting delay chain to zero\n",
2237 __func__, __LINE__, rw_group, !ret);
2239 for (r = 0; r < RW_MGR_MEM_NUMBER_OF_RANKS;
2240 r += NUM_RANKS_PER_SHADOW_REG) {
2241 scc_mgr_apply_group_dq_in_delay(test_bgn, 0);
2242 writel(0, &sdr_scc_mgr->update);
2249 * rw_mgr_mem_calibrate_dq_dqs_centering() - Centering DQ/DQS
2250 * @rw_group: Read/Write Group
2251 * @test_bgn: Rank at which the test begins
2252 * @use_read_test: Perform a read test
2253 * @update_fom: Update FOM
2255 * The centerin DQ/DQS stage attempts to align DQ and DQS signals on reads
2259 rw_mgr_mem_calibrate_dq_dqs_centering(const u32 rw_group, const u32 test_bgn,
2260 const int use_read_test,
2261 const int update_fom)
2264 int ret, grp_calibrated;
2268 * Altera EMI_RM 2015.05.04 :: Figure 1-28
2269 * Read per-bit deskew can be done on a per shadow register basis.
2272 for (rank_bgn = 0, sr = 0;
2273 rank_bgn < RW_MGR_MEM_NUMBER_OF_RANKS;
2274 rank_bgn += NUM_RANKS_PER_SHADOW_REG, sr++) {
2275 /* Check if this set of ranks should be skipped entirely. */
2276 if (param->skip_shadow_regs[sr])
2279 ret = rw_mgr_mem_calibrate_vfifo_center(rank_bgn, rw_group,
2289 if (!grp_calibrated)
2296 * rw_mgr_mem_calibrate_vfifo() - Calibrate the read valid prediction FIFO
2297 * @rw_group: Read/Write Group
2298 * @test_bgn: Rank at which the test begins
2300 * Stage 1: Calibrate the read valid prediction FIFO.
2302 * This function implements UniPHY calibration Stage 1, as explained in
2303 * detail in Altera EMI_RM 2015.05.04 , "UniPHY Calibration Stages".
2305 * - read valid prediction will consist of finding:
2306 * - DQS enable phase and DQS enable delay (DQS Enable Calibration)
2307 * - DQS input phase and DQS input delay (DQ/DQS Centering)
2308 * - we also do a per-bit deskew on the DQ lines.
2310 static int rw_mgr_mem_calibrate_vfifo(const u32 rw_group, const u32 test_bgn)
2313 uint32_t dtaps_per_ptap;
2314 uint32_t failed_substage;
2318 debug("%s:%d: %u %u\n", __func__, __LINE__, rw_group, test_bgn);
2320 /* Update info for sims */
2321 reg_file_set_group(rw_group);
2322 reg_file_set_stage(CAL_STAGE_VFIFO);
2323 reg_file_set_sub_stage(CAL_SUBSTAGE_GUARANTEED_READ);
2325 failed_substage = CAL_SUBSTAGE_GUARANTEED_READ;
2327 /* USER Determine number of delay taps for each phase tap. */
2328 dtaps_per_ptap = DIV_ROUND_UP(IO_DELAY_PER_OPA_TAP,
2329 IO_DELAY_PER_DQS_EN_DCHAIN_TAP) - 1;
2331 for (d = 0; d <= dtaps_per_ptap; d += 2) {
2333 * In RLDRAMX we may be messing the delay of pins in
2334 * the same write rw_group but outside of the current read
2335 * the rw_group, but that's ok because we haven't calibrated
2339 scc_mgr_apply_group_all_out_delay_add_all_ranks(
2343 for (p = 0; p <= IO_DQDQS_OUT_PHASE_MAX; p++) {
2344 /* 1) Guaranteed Write */
2345 ret = rw_mgr_mem_calibrate_guaranteed_write(rw_group, p);
2349 /* 2) DQS Enable Calibration */
2350 ret = rw_mgr_mem_calibrate_dqs_enable_calibration(rw_group,
2353 failed_substage = CAL_SUBSTAGE_DQS_EN_PHASE;
2357 /* 3) Centering DQ/DQS */
2359 * If doing read after write calibration, do not update
2360 * FOM now. Do it then.
2362 ret = rw_mgr_mem_calibrate_dq_dqs_centering(rw_group,
2365 failed_substage = CAL_SUBSTAGE_VFIFO_CENTER;
2374 /* Calibration Stage 1 failed. */
2375 set_failing_group_stage(rw_group, CAL_STAGE_VFIFO, failed_substage);
2378 /* Calibration Stage 1 completed OK. */
2381 * Reset the delay chains back to zero if they have moved > 1
2382 * (check for > 1 because loop will increase d even when pass in
2386 scc_mgr_zero_group(rw_group, 1);
2391 /* VFIFO Calibration -- Read Deskew Calibration after write deskew */
2392 static uint32_t rw_mgr_mem_calibrate_vfifo_end(uint32_t read_group,
2395 uint32_t rank_bgn, sr;
2396 uint32_t grp_calibrated;
2397 uint32_t write_group;
2399 debug("%s:%d %u %u", __func__, __LINE__, read_group, test_bgn);
2401 /* update info for sims */
2403 reg_file_set_stage(CAL_STAGE_VFIFO_AFTER_WRITES);
2404 reg_file_set_sub_stage(CAL_SUBSTAGE_VFIFO_CENTER);
2406 write_group = read_group;
2408 /* update info for sims */
2409 reg_file_set_group(read_group);
2412 /* Read per-bit deskew can be done on a per shadow register basis */
2413 for (rank_bgn = 0, sr = 0; rank_bgn < RW_MGR_MEM_NUMBER_OF_RANKS;
2414 rank_bgn += NUM_RANKS_PER_SHADOW_REG, ++sr) {
2415 /* Determine if this set of ranks should be skipped entirely */
2416 if (!param->skip_shadow_regs[sr]) {
2417 /* This is the last calibration round, update FOM here */
2418 if (!rw_mgr_mem_calibrate_vfifo_center(rank_bgn,
2429 if (grp_calibrated == 0) {
2430 set_failing_group_stage(write_group,
2431 CAL_STAGE_VFIFO_AFTER_WRITES,
2432 CAL_SUBSTAGE_VFIFO_CENTER);
2439 /* Calibrate LFIFO to find smallest read latency */
2440 static uint32_t rw_mgr_mem_calibrate_lfifo(void)
2444 debug("%s:%d\n", __func__, __LINE__);
2446 /* update info for sims */
2447 reg_file_set_stage(CAL_STAGE_LFIFO);
2448 reg_file_set_sub_stage(CAL_SUBSTAGE_READ_LATENCY);
2450 /* Load up the patterns used by read calibration for all ranks */
2451 rw_mgr_mem_calibrate_read_load_patterns(0, 1);
2455 writel(gbl->curr_read_lat, &phy_mgr_cfg->phy_rlat);
2456 debug_cond(DLEVEL == 2, "%s:%d lfifo: read_lat=%u",
2457 __func__, __LINE__, gbl->curr_read_lat);
2459 if (!rw_mgr_mem_calibrate_read_test_all_ranks(0,
2467 /* reduce read latency and see if things are working */
2469 gbl->curr_read_lat--;
2470 } while (gbl->curr_read_lat > 0);
2472 /* reset the fifos to get pointers to known state */
2474 writel(0, &phy_mgr_cmd->fifo_reset);
2477 /* add a fudge factor to the read latency that was determined */
2478 gbl->curr_read_lat += 2;
2479 writel(gbl->curr_read_lat, &phy_mgr_cfg->phy_rlat);
2480 debug_cond(DLEVEL == 2, "%s:%d lfifo: success: using \
2481 read_lat=%u\n", __func__, __LINE__,
2482 gbl->curr_read_lat);
2485 set_failing_group_stage(0xff, CAL_STAGE_LFIFO,
2486 CAL_SUBSTAGE_READ_LATENCY);
2488 debug_cond(DLEVEL == 2, "%s:%d lfifo: failed at initial \
2489 read_lat=%u\n", __func__, __LINE__,
2490 gbl->curr_read_lat);
2496 * issue write test command.
2497 * two variants are provided. one that just tests a write pattern and
2498 * another that tests datamask functionality.
2500 static void rw_mgr_mem_calibrate_write_test_issue(uint32_t group,
2503 uint32_t mcc_instruction;
2504 uint32_t quick_write_mode = (((STATIC_CALIB_STEPS) & CALIB_SKIP_WRITES) &&
2505 ENABLE_SUPER_QUICK_CALIBRATION);
2506 uint32_t rw_wl_nop_cycles;
2510 * Set counter and jump addresses for the right
2511 * number of NOP cycles.
2512 * The number of supported NOP cycles can range from -1 to infinity
2513 * Three different cases are handled:
2515 * 1. For a number of NOP cycles greater than 0, the RW Mgr looping
2516 * mechanism will be used to insert the right number of NOPs
2518 * 2. For a number of NOP cycles equals to 0, the micro-instruction
2519 * issuing the write command will jump straight to the
2520 * micro-instruction that turns on DQS (for DDRx), or outputs write
2521 * data (for RLD), skipping
2522 * the NOP micro-instruction all together
2524 * 3. A number of NOP cycles equal to -1 indicates that DQS must be
2525 * turned on in the same micro-instruction that issues the write
2526 * command. Then we need
2527 * to directly jump to the micro-instruction that sends out the data
2529 * NOTE: Implementing this mechanism uses 2 RW Mgr jump-counters
2530 * (2 and 3). One jump-counter (0) is used to perform multiple
2531 * write-read operations.
2532 * one counter left to issue this command in "multiple-group" mode
2535 rw_wl_nop_cycles = gbl->rw_wl_nop_cycles;
2537 if (rw_wl_nop_cycles == -1) {
2539 * CNTR 2 - We want to execute the special write operation that
2540 * turns on DQS right away and then skip directly to the
2541 * instruction that sends out the data. We set the counter to a
2542 * large number so that the jump is always taken.
2544 writel(0xFF, &sdr_rw_load_mgr_regs->load_cntr2);
2546 /* CNTR 3 - Not used */
2548 mcc_instruction = RW_MGR_LFSR_WR_RD_DM_BANK_0_WL_1;
2549 writel(RW_MGR_LFSR_WR_RD_DM_BANK_0_DATA,
2550 &sdr_rw_load_jump_mgr_regs->load_jump_add2);
2551 writel(RW_MGR_LFSR_WR_RD_DM_BANK_0_NOP,
2552 &sdr_rw_load_jump_mgr_regs->load_jump_add3);
2554 mcc_instruction = RW_MGR_LFSR_WR_RD_BANK_0_WL_1;
2555 writel(RW_MGR_LFSR_WR_RD_BANK_0_DATA,
2556 &sdr_rw_load_jump_mgr_regs->load_jump_add2);
2557 writel(RW_MGR_LFSR_WR_RD_BANK_0_NOP,
2558 &sdr_rw_load_jump_mgr_regs->load_jump_add3);
2560 } else if (rw_wl_nop_cycles == 0) {
2562 * CNTR 2 - We want to skip the NOP operation and go straight
2563 * to the DQS enable instruction. We set the counter to a large
2564 * number so that the jump is always taken.
2566 writel(0xFF, &sdr_rw_load_mgr_regs->load_cntr2);
2568 /* CNTR 3 - Not used */
2570 mcc_instruction = RW_MGR_LFSR_WR_RD_DM_BANK_0;
2571 writel(RW_MGR_LFSR_WR_RD_DM_BANK_0_DQS,
2572 &sdr_rw_load_jump_mgr_regs->load_jump_add2);
2574 mcc_instruction = RW_MGR_LFSR_WR_RD_BANK_0;
2575 writel(RW_MGR_LFSR_WR_RD_BANK_0_DQS,
2576 &sdr_rw_load_jump_mgr_regs->load_jump_add2);
2580 * CNTR 2 - In this case we want to execute the next instruction
2581 * and NOT take the jump. So we set the counter to 0. The jump
2582 * address doesn't count.
2584 writel(0x0, &sdr_rw_load_mgr_regs->load_cntr2);
2585 writel(0x0, &sdr_rw_load_jump_mgr_regs->load_jump_add2);
2588 * CNTR 3 - Set the nop counter to the number of cycles we
2589 * need to loop for, minus 1.
2591 writel(rw_wl_nop_cycles - 1, &sdr_rw_load_mgr_regs->load_cntr3);
2593 mcc_instruction = RW_MGR_LFSR_WR_RD_DM_BANK_0;
2594 writel(RW_MGR_LFSR_WR_RD_DM_BANK_0_NOP,
2595 &sdr_rw_load_jump_mgr_regs->load_jump_add3);
2597 mcc_instruction = RW_MGR_LFSR_WR_RD_BANK_0;
2598 writel(RW_MGR_LFSR_WR_RD_BANK_0_NOP,
2599 &sdr_rw_load_jump_mgr_regs->load_jump_add3);
2603 writel(0, SDR_PHYGRP_RWMGRGRP_ADDRESS |
2604 RW_MGR_RESET_READ_DATAPATH_OFFSET);
2606 if (quick_write_mode)
2607 writel(0x08, &sdr_rw_load_mgr_regs->load_cntr0);
2609 writel(0x40, &sdr_rw_load_mgr_regs->load_cntr0);
2611 writel(mcc_instruction, &sdr_rw_load_jump_mgr_regs->load_jump_add0);
2614 * CNTR 1 - This is used to ensure enough time elapses
2615 * for read data to come back.
2617 writel(0x30, &sdr_rw_load_mgr_regs->load_cntr1);
2620 writel(RW_MGR_LFSR_WR_RD_DM_BANK_0_WAIT,
2621 &sdr_rw_load_jump_mgr_regs->load_jump_add1);
2623 writel(RW_MGR_LFSR_WR_RD_BANK_0_WAIT,
2624 &sdr_rw_load_jump_mgr_regs->load_jump_add1);
2627 addr = SDR_PHYGRP_RWMGRGRP_ADDRESS | RW_MGR_RUN_SINGLE_GROUP_OFFSET;
2628 writel(mcc_instruction, addr + (group << 2));
2631 /* Test writes, can check for a single bit pass or multiple bit pass */
2632 static uint32_t rw_mgr_mem_calibrate_write_test(uint32_t rank_bgn,
2633 uint32_t write_group, uint32_t use_dm, uint32_t all_correct,
2634 uint32_t *bit_chk, uint32_t all_ranks)
2637 uint32_t correct_mask_vg;
2638 uint32_t tmp_bit_chk;
2640 uint32_t rank_end = all_ranks ? RW_MGR_MEM_NUMBER_OF_RANKS :
2641 (rank_bgn + NUM_RANKS_PER_SHADOW_REG);
2642 uint32_t addr_rw_mgr;
2643 uint32_t base_rw_mgr;
2645 *bit_chk = param->write_correct_mask;
2646 correct_mask_vg = param->write_correct_mask_vg;
2648 for (r = rank_bgn; r < rank_end; r++) {
2649 if (param->skip_ranks[r]) {
2650 /* request to skip the rank */
2655 set_rank_and_odt_mask(r, RW_MGR_ODT_MODE_READ_WRITE);
2658 addr_rw_mgr = SDR_PHYGRP_RWMGRGRP_ADDRESS;
2659 for (vg = RW_MGR_MEM_VIRTUAL_GROUPS_PER_WRITE_DQS-1; ; vg--) {
2660 /* reset the fifos to get pointers to known state */
2661 writel(0, &phy_mgr_cmd->fifo_reset);
2663 tmp_bit_chk = tmp_bit_chk <<
2664 (RW_MGR_MEM_DQ_PER_WRITE_DQS /
2665 RW_MGR_MEM_VIRTUAL_GROUPS_PER_WRITE_DQS);
2666 rw_mgr_mem_calibrate_write_test_issue(write_group *
2667 RW_MGR_MEM_VIRTUAL_GROUPS_PER_WRITE_DQS+vg,
2670 base_rw_mgr = readl(addr_rw_mgr);
2671 tmp_bit_chk = tmp_bit_chk | (correct_mask_vg & ~(base_rw_mgr));
2675 *bit_chk &= tmp_bit_chk;
2679 set_rank_and_odt_mask(0, RW_MGR_ODT_MODE_OFF);
2680 debug_cond(DLEVEL == 2, "write_test(%u,%u,ALL) : %u == \
2681 %u => %lu", write_group, use_dm,
2682 *bit_chk, param->write_correct_mask,
2683 (long unsigned int)(*bit_chk ==
2684 param->write_correct_mask));
2685 return *bit_chk == param->write_correct_mask;
2687 set_rank_and_odt_mask(0, RW_MGR_ODT_MODE_OFF);
2688 debug_cond(DLEVEL == 2, "write_test(%u,%u,ONE) : %u != ",
2689 write_group, use_dm, *bit_chk);
2690 debug_cond(DLEVEL == 2, "%lu" " => %lu", (long unsigned int)0,
2691 (long unsigned int)(*bit_chk != 0));
2692 return *bit_chk != 0x00;
2697 * center all windows. do per-bit-deskew to possibly increase size of
2700 static uint32_t rw_mgr_mem_calibrate_writes_center(uint32_t rank_bgn,
2701 uint32_t write_group, uint32_t test_bgn)
2703 uint32_t i, p, min_index;
2706 * Store these as signed since there are comparisons with
2710 uint32_t sticky_bit_chk;
2711 int32_t left_edge[RW_MGR_MEM_DQ_PER_WRITE_DQS];
2712 int32_t right_edge[RW_MGR_MEM_DQ_PER_WRITE_DQS];
2714 int32_t mid_min, orig_mid_min;
2715 int32_t new_dqs, start_dqs, shift_dq;
2716 int32_t dq_margin, dqs_margin, dm_margin;
2718 uint32_t temp_dq_out1_delay;
2721 debug("%s:%d %u %u", __func__, __LINE__, write_group, test_bgn);
2725 addr = SDR_PHYGRP_SCCGRP_ADDRESS | SCC_MGR_IO_OUT1_DELAY_OFFSET;
2726 start_dqs = readl(addr +
2727 (RW_MGR_MEM_DQ_PER_WRITE_DQS << 2));
2729 /* per-bit deskew */
2732 * set the left and right edge of each bit to an illegal value
2733 * use (IO_IO_OUT1_DELAY_MAX + 1) as an illegal value.
2736 for (i = 0; i < RW_MGR_MEM_DQ_PER_WRITE_DQS; i++) {
2737 left_edge[i] = IO_IO_OUT1_DELAY_MAX + 1;
2738 right_edge[i] = IO_IO_OUT1_DELAY_MAX + 1;
2741 /* Search for the left edge of the window for each bit */
2742 for (d = 0; d <= IO_IO_OUT1_DELAY_MAX; d++) {
2743 scc_mgr_apply_group_dq_out1_delay(write_group, d);
2745 writel(0, &sdr_scc_mgr->update);
2748 * Stop searching when the read test doesn't pass AND when
2749 * we've seen a passing read on every bit.
2751 stop = !rw_mgr_mem_calibrate_write_test(rank_bgn, write_group,
2752 0, PASS_ONE_BIT, &bit_chk, 0);
2753 sticky_bit_chk = sticky_bit_chk | bit_chk;
2754 stop = stop && (sticky_bit_chk == param->write_correct_mask);
2755 debug_cond(DLEVEL == 2, "write_center(left): dtap=%d => %u \
2756 == %u && %u [bit_chk= %u ]\n",
2757 d, sticky_bit_chk, param->write_correct_mask,
2763 for (i = 0; i < RW_MGR_MEM_DQ_PER_WRITE_DQS; i++) {
2766 * Remember a passing test as the
2772 * If a left edge has not been seen
2773 * yet, then a future passing test will
2774 * mark this edge as the right edge.
2777 IO_IO_OUT1_DELAY_MAX + 1) {
2778 right_edge[i] = -(d + 1);
2781 debug_cond(DLEVEL == 2, "write_center[l,d=%d):", d);
2782 debug_cond(DLEVEL == 2, "bit_chk_test=%d left_edge[%u]: %d",
2783 (int)(bit_chk & 1), i, left_edge[i]);
2784 debug_cond(DLEVEL == 2, "right_edge[%u]: %d\n", i,
2786 bit_chk = bit_chk >> 1;
2791 /* Reset DQ delay chains to 0 */
2792 scc_mgr_apply_group_dq_out1_delay(0);
2794 for (i = RW_MGR_MEM_DQ_PER_WRITE_DQS - 1;; i--) {
2795 debug_cond(DLEVEL == 2, "%s:%d write_center: left_edge[%u]: \
2796 %d right_edge[%u]: %d\n", __func__, __LINE__,
2797 i, left_edge[i], i, right_edge[i]);
2800 * Check for cases where we haven't found the left edge,
2801 * which makes our assignment of the the right edge invalid.
2802 * Reset it to the illegal value.
2804 if ((left_edge[i] == IO_IO_OUT1_DELAY_MAX + 1) &&
2805 (right_edge[i] != IO_IO_OUT1_DELAY_MAX + 1)) {
2806 right_edge[i] = IO_IO_OUT1_DELAY_MAX + 1;
2807 debug_cond(DLEVEL == 2, "%s:%d write_center: reset \
2808 right_edge[%u]: %d\n", __func__, __LINE__,
2813 * Reset sticky bit (except for bits where we have
2814 * seen the left edge).
2816 sticky_bit_chk = sticky_bit_chk << 1;
2817 if ((left_edge[i] != IO_IO_OUT1_DELAY_MAX + 1))
2818 sticky_bit_chk = sticky_bit_chk | 1;
2824 /* Search for the right edge of the window for each bit */
2825 for (d = 0; d <= IO_IO_OUT1_DELAY_MAX - start_dqs; d++) {
2826 scc_mgr_apply_group_dqs_io_and_oct_out1(write_group,
2829 writel(0, &sdr_scc_mgr->update);
2832 * Stop searching when the read test doesn't pass AND when
2833 * we've seen a passing read on every bit.
2835 stop = !rw_mgr_mem_calibrate_write_test(rank_bgn, write_group,
2836 0, PASS_ONE_BIT, &bit_chk, 0);
2838 sticky_bit_chk = sticky_bit_chk | bit_chk;
2839 stop = stop && (sticky_bit_chk == param->write_correct_mask);
2841 debug_cond(DLEVEL == 2, "write_center (right): dtap=%u => %u == \
2842 %u && %u\n", d, sticky_bit_chk,
2843 param->write_correct_mask, stop);
2847 for (i = 0; i < RW_MGR_MEM_DQ_PER_WRITE_DQS;
2849 /* d = 0 failed, but it passed when
2850 testing the left edge, so it must be
2851 marginal, set it to -1 */
2852 if (right_edge[i] ==
2853 IO_IO_OUT1_DELAY_MAX + 1 &&
2855 IO_IO_OUT1_DELAY_MAX + 1) {
2862 for (i = 0; i < RW_MGR_MEM_DQ_PER_WRITE_DQS; i++) {
2865 * Remember a passing test as
2872 * If a right edge has not
2873 * been seen yet, then a future
2874 * passing test will mark this
2875 * edge as the left edge.
2877 if (right_edge[i] ==
2878 IO_IO_OUT1_DELAY_MAX + 1)
2879 left_edge[i] = -(d + 1);
2882 * d = 0 failed, but it passed
2883 * when testing the left edge,
2884 * so it must be marginal, set
2887 if (right_edge[i] ==
2888 IO_IO_OUT1_DELAY_MAX + 1 &&
2890 IO_IO_OUT1_DELAY_MAX + 1)
2893 * If a right edge has not been
2894 * seen yet, then a future
2895 * passing test will mark this
2896 * edge as the left edge.
2898 else if (right_edge[i] ==
2899 IO_IO_OUT1_DELAY_MAX +
2901 left_edge[i] = -(d + 1);
2904 debug_cond(DLEVEL == 2, "write_center[r,d=%d):", d);
2905 debug_cond(DLEVEL == 2, "bit_chk_test=%d left_edge[%u]: %d",
2906 (int)(bit_chk & 1), i, left_edge[i]);
2907 debug_cond(DLEVEL == 2, "right_edge[%u]: %d\n", i,
2909 bit_chk = bit_chk >> 1;
2914 /* Check that all bits have a window */
2915 for (i = 0; i < RW_MGR_MEM_DQ_PER_WRITE_DQS; i++) {
2916 debug_cond(DLEVEL == 2, "%s:%d write_center: left_edge[%u]: \
2917 %d right_edge[%u]: %d", __func__, __LINE__,
2918 i, left_edge[i], i, right_edge[i]);
2919 if ((left_edge[i] == IO_IO_OUT1_DELAY_MAX + 1) ||
2920 (right_edge[i] == IO_IO_OUT1_DELAY_MAX + 1)) {
2921 set_failing_group_stage(test_bgn + i,
2923 CAL_SUBSTAGE_WRITES_CENTER);
2928 /* Find middle of window for each DQ bit */
2929 mid_min = left_edge[0] - right_edge[0];
2931 for (i = 1; i < RW_MGR_MEM_DQ_PER_WRITE_DQS; i++) {
2932 mid = left_edge[i] - right_edge[i];
2933 if (mid < mid_min) {
2940 * -mid_min/2 represents the amount that we need to move DQS.
2941 * If mid_min is odd and positive we'll need to add one to
2942 * make sure the rounding in further calculations is correct
2943 * (always bias to the right), so just add 1 for all positive values.
2947 mid_min = mid_min / 2;
2948 debug_cond(DLEVEL == 1, "%s:%d write_center: mid_min=%d\n", __func__,
2951 /* Determine the amount we can change DQS (which is -mid_min) */
2952 orig_mid_min = mid_min;
2953 new_dqs = start_dqs;
2955 debug_cond(DLEVEL == 1, "%s:%d write_center: start_dqs=%d new_dqs=%d \
2956 mid_min=%d\n", __func__, __LINE__, start_dqs, new_dqs, mid_min);
2957 /* Initialize data for export structures */
2958 dqs_margin = IO_IO_OUT1_DELAY_MAX + 1;
2959 dq_margin = IO_IO_OUT1_DELAY_MAX + 1;
2961 /* add delay to bring centre of all DQ windows to the same "level" */
2962 for (i = 0, p = test_bgn; i < RW_MGR_MEM_DQ_PER_WRITE_DQS; i++, p++) {
2963 /* Use values before divide by 2 to reduce round off error */
2964 shift_dq = (left_edge[i] - right_edge[i] -
2965 (left_edge[min_index] - right_edge[min_index]))/2 +
2966 (orig_mid_min - mid_min);
2968 debug_cond(DLEVEL == 2, "%s:%d write_center: before: shift_dq \
2969 [%u]=%d\n", __func__, __LINE__, i, shift_dq);
2971 addr = SDR_PHYGRP_SCCGRP_ADDRESS | SCC_MGR_IO_OUT1_DELAY_OFFSET;
2972 temp_dq_out1_delay = readl(addr + (i << 2));
2973 if (shift_dq + (int32_t)temp_dq_out1_delay >
2974 (int32_t)IO_IO_OUT1_DELAY_MAX) {
2975 shift_dq = (int32_t)IO_IO_OUT1_DELAY_MAX - temp_dq_out1_delay;
2976 } else if (shift_dq + (int32_t)temp_dq_out1_delay < 0) {
2977 shift_dq = -(int32_t)temp_dq_out1_delay;
2979 debug_cond(DLEVEL == 2, "write_center: after: shift_dq[%u]=%d\n",
2981 scc_mgr_set_dq_out1_delay(i, temp_dq_out1_delay + shift_dq);
2984 debug_cond(DLEVEL == 2, "write_center: margin[%u]=[%d,%d]\n", i,
2985 left_edge[i] - shift_dq + (-mid_min),
2986 right_edge[i] + shift_dq - (-mid_min));
2987 /* To determine values for export structures */
2988 if (left_edge[i] - shift_dq + (-mid_min) < dq_margin)
2989 dq_margin = left_edge[i] - shift_dq + (-mid_min);
2991 if (right_edge[i] + shift_dq - (-mid_min) < dqs_margin)
2992 dqs_margin = right_edge[i] + shift_dq - (-mid_min);
2996 scc_mgr_apply_group_dqs_io_and_oct_out1(write_group, new_dqs);
2997 writel(0, &sdr_scc_mgr->update);
3000 debug_cond(DLEVEL == 2, "%s:%d write_center: DM\n", __func__, __LINE__);
3003 * set the left and right edge of each bit to an illegal value,
3004 * use (IO_IO_OUT1_DELAY_MAX + 1) as an illegal value,
3006 left_edge[0] = IO_IO_OUT1_DELAY_MAX + 1;
3007 right_edge[0] = IO_IO_OUT1_DELAY_MAX + 1;
3008 int32_t bgn_curr = IO_IO_OUT1_DELAY_MAX + 1;
3009 int32_t end_curr = IO_IO_OUT1_DELAY_MAX + 1;
3010 int32_t bgn_best = IO_IO_OUT1_DELAY_MAX + 1;
3011 int32_t end_best = IO_IO_OUT1_DELAY_MAX + 1;
3012 int32_t win_best = 0;
3014 /* Search for the/part of the window with DM shift */
3015 for (d = IO_IO_OUT1_DELAY_MAX; d >= 0; d -= DELTA_D) {
3016 scc_mgr_apply_group_dm_out1_delay(d);
3017 writel(0, &sdr_scc_mgr->update);
3019 if (rw_mgr_mem_calibrate_write_test(rank_bgn, write_group, 1,
3020 PASS_ALL_BITS, &bit_chk,
3022 /* USE Set current end of the window */
3025 * If a starting edge of our window has not been seen
3026 * this is our current start of the DM window.
3028 if (bgn_curr == IO_IO_OUT1_DELAY_MAX + 1)
3032 * If current window is bigger than best seen.
3033 * Set best seen to be current window.
3035 if ((end_curr-bgn_curr+1) > win_best) {
3036 win_best = end_curr-bgn_curr+1;
3037 bgn_best = bgn_curr;
3038 end_best = end_curr;
3041 /* We just saw a failing test. Reset temp edge */
3042 bgn_curr = IO_IO_OUT1_DELAY_MAX + 1;
3043 end_curr = IO_IO_OUT1_DELAY_MAX + 1;
3048 /* Reset DM delay chains to 0 */
3049 scc_mgr_apply_group_dm_out1_delay(0);
3052 * Check to see if the current window nudges up aganist 0 delay.
3053 * If so we need to continue the search by shifting DQS otherwise DQS
3054 * search begins as a new search. */
3055 if (end_curr != 0) {
3056 bgn_curr = IO_IO_OUT1_DELAY_MAX + 1;
3057 end_curr = IO_IO_OUT1_DELAY_MAX + 1;
3060 /* Search for the/part of the window with DQS shifts */
3061 for (d = 0; d <= IO_IO_OUT1_DELAY_MAX - new_dqs; d += DELTA_D) {
3063 * Note: This only shifts DQS, so are we limiting ourselve to
3064 * width of DQ unnecessarily.
3066 scc_mgr_apply_group_dqs_io_and_oct_out1(write_group,
3069 writel(0, &sdr_scc_mgr->update);
3070 if (rw_mgr_mem_calibrate_write_test(rank_bgn, write_group, 1,
3071 PASS_ALL_BITS, &bit_chk,
3073 /* USE Set current end of the window */
3076 * If a beginning edge of our window has not been seen
3077 * this is our current begin of the DM window.
3079 if (bgn_curr == IO_IO_OUT1_DELAY_MAX + 1)
3083 * If current window is bigger than best seen. Set best
3084 * seen to be current window.
3086 if ((end_curr-bgn_curr+1) > win_best) {
3087 win_best = end_curr-bgn_curr+1;
3088 bgn_best = bgn_curr;
3089 end_best = end_curr;
3092 /* We just saw a failing test. Reset temp edge */
3093 bgn_curr = IO_IO_OUT1_DELAY_MAX + 1;
3094 end_curr = IO_IO_OUT1_DELAY_MAX + 1;
3096 /* Early exit optimization: if ther remaining delay
3097 chain space is less than already seen largest window
3100 (IO_IO_OUT1_DELAY_MAX - new_dqs - d)) {
3106 /* assign left and right edge for cal and reporting; */
3107 left_edge[0] = -1*bgn_best;
3108 right_edge[0] = end_best;
3110 debug_cond(DLEVEL == 2, "%s:%d dm_calib: left=%d right=%d\n", __func__,
3111 __LINE__, left_edge[0], right_edge[0]);
3113 /* Move DQS (back to orig) */
3114 scc_mgr_apply_group_dqs_io_and_oct_out1(write_group, new_dqs);
3118 /* Find middle of window for the DM bit */
3119 mid = (left_edge[0] - right_edge[0]) / 2;
3121 /* only move right, since we are not moving DQS/DQ */
3125 /* dm_marign should fail if we never find a window */
3129 dm_margin = left_edge[0] - mid;
3131 scc_mgr_apply_group_dm_out1_delay(mid);
3132 writel(0, &sdr_scc_mgr->update);
3134 debug_cond(DLEVEL == 2, "%s:%d dm_calib: left=%d right=%d mid=%d \
3135 dm_margin=%d\n", __func__, __LINE__, left_edge[0],
3136 right_edge[0], mid, dm_margin);
3138 gbl->fom_out += dq_margin + dqs_margin;
3140 debug_cond(DLEVEL == 2, "%s:%d write_center: dq_margin=%d \
3141 dqs_margin=%d dm_margin=%d\n", __func__, __LINE__,
3142 dq_margin, dqs_margin, dm_margin);
3145 * Do not remove this line as it makes sure all of our
3146 * decisions have been applied.
3148 writel(0, &sdr_scc_mgr->update);
3149 return (dq_margin >= 0) && (dqs_margin >= 0) && (dm_margin >= 0);
3152 /* calibrate the write operations */
3153 static uint32_t rw_mgr_mem_calibrate_writes(uint32_t rank_bgn, uint32_t g,
3156 /* update info for sims */
3157 debug("%s:%d %u %u\n", __func__, __LINE__, g, test_bgn);
3159 reg_file_set_stage(CAL_STAGE_WRITES);
3160 reg_file_set_sub_stage(CAL_SUBSTAGE_WRITES_CENTER);
3162 reg_file_set_group(g);
3164 if (!rw_mgr_mem_calibrate_writes_center(rank_bgn, g, test_bgn)) {
3165 set_failing_group_stage(g, CAL_STAGE_WRITES,
3166 CAL_SUBSTAGE_WRITES_CENTER);
3174 * mem_precharge_and_activate() - Precharge all banks and activate
3176 * Precharge all banks and activate row 0 in bank "000..." and bank "111...".
3178 static void mem_precharge_and_activate(void)
3182 for (r = 0; r < RW_MGR_MEM_NUMBER_OF_RANKS; r++) {
3183 /* Test if the rank should be skipped. */
3184 if (param->skip_ranks[r])
3188 set_rank_and_odt_mask(r, RW_MGR_ODT_MODE_OFF);
3190 /* Precharge all banks. */
3191 writel(RW_MGR_PRECHARGE_ALL, SDR_PHYGRP_RWMGRGRP_ADDRESS |
3192 RW_MGR_RUN_SINGLE_GROUP_OFFSET);
3194 writel(0x0F, &sdr_rw_load_mgr_regs->load_cntr0);
3195 writel(RW_MGR_ACTIVATE_0_AND_1_WAIT1,
3196 &sdr_rw_load_jump_mgr_regs->load_jump_add0);
3198 writel(0x0F, &sdr_rw_load_mgr_regs->load_cntr1);
3199 writel(RW_MGR_ACTIVATE_0_AND_1_WAIT2,
3200 &sdr_rw_load_jump_mgr_regs->load_jump_add1);
3202 /* Activate rows. */
3203 writel(RW_MGR_ACTIVATE_0_AND_1, SDR_PHYGRP_RWMGRGRP_ADDRESS |
3204 RW_MGR_RUN_SINGLE_GROUP_OFFSET);
3209 * mem_init_latency() - Configure memory RLAT and WLAT settings
3211 * Configure memory RLAT and WLAT parameters.
3213 static void mem_init_latency(void)
3216 * For AV/CV, LFIFO is hardened and always runs at full rate
3217 * so max latency in AFI clocks, used here, is correspondingly
3220 const u32 max_latency = (1 << MAX_LATENCY_COUNT_WIDTH) - 1;
3223 debug("%s:%d\n", __func__, __LINE__);
3226 * Read in write latency.
3227 * WL for Hard PHY does not include additive latency.
3229 wlat = readl(&data_mgr->t_wl_add);
3230 wlat += readl(&data_mgr->mem_t_add);
3232 gbl->rw_wl_nop_cycles = wlat - 1;
3234 /* Read in readl latency. */
3235 rlat = readl(&data_mgr->t_rl_add);
3237 /* Set a pretty high read latency initially. */
3238 gbl->curr_read_lat = rlat + 16;
3239 if (gbl->curr_read_lat > max_latency)
3240 gbl->curr_read_lat = max_latency;
3242 writel(gbl->curr_read_lat, &phy_mgr_cfg->phy_rlat);
3244 /* Advertise write latency. */
3245 writel(wlat, &phy_mgr_cfg->afi_wlat);
3249 * @mem_skip_calibrate() - Set VFIFO and LFIFO to instant-on settings
3251 * Set VFIFO and LFIFO to instant-on settings in skip calibration mode.
3253 static void mem_skip_calibrate(void)
3255 uint32_t vfifo_offset;
3258 debug("%s:%d\n", __func__, __LINE__);
3259 /* Need to update every shadow register set used by the interface */
3260 for (r = 0; r < RW_MGR_MEM_NUMBER_OF_RANKS;
3261 r += NUM_RANKS_PER_SHADOW_REG) {
3263 * Set output phase alignment settings appropriate for
3266 for (i = 0; i < RW_MGR_MEM_IF_READ_DQS_WIDTH; i++) {
3267 scc_mgr_set_dqs_en_phase(i, 0);
3268 #if IO_DLL_CHAIN_LENGTH == 6
3269 scc_mgr_set_dqdqs_output_phase(i, 6);
3271 scc_mgr_set_dqdqs_output_phase(i, 7);
3276 * Write data arrives to the I/O two cycles before write
3277 * latency is reached (720 deg).
3278 * -> due to bit-slip in a/c bus
3279 * -> to allow board skew where dqs is longer than ck
3280 * -> how often can this happen!?
3281 * -> can claim back some ptaps for high freq
3282 * support if we can relax this, but i digress...
3284 * The write_clk leads mem_ck by 90 deg
3285 * The minimum ptap of the OPA is 180 deg
3286 * Each ptap has (360 / IO_DLL_CHAIN_LENGH) deg of delay
3287 * The write_clk is always delayed by 2 ptaps
3289 * Hence, to make DQS aligned to CK, we need to delay
3291 * (720 - 90 - 180 - 2 * (360 / IO_DLL_CHAIN_LENGTH))
3293 * Dividing the above by (360 / IO_DLL_CHAIN_LENGTH)
3294 * gives us the number of ptaps, which simplies to:
3296 * (1.25 * IO_DLL_CHAIN_LENGTH - 2)
3298 scc_mgr_set_dqdqs_output_phase(i,
3299 1.25 * IO_DLL_CHAIN_LENGTH - 2);
3301 writel(0xff, &sdr_scc_mgr->dqs_ena);
3302 writel(0xff, &sdr_scc_mgr->dqs_io_ena);
3304 for (i = 0; i < RW_MGR_MEM_IF_WRITE_DQS_WIDTH; i++) {
3305 writel(i, SDR_PHYGRP_SCCGRP_ADDRESS |
3306 SCC_MGR_GROUP_COUNTER_OFFSET);
3308 writel(0xff, &sdr_scc_mgr->dq_ena);
3309 writel(0xff, &sdr_scc_mgr->dm_ena);
3310 writel(0, &sdr_scc_mgr->update);
3313 /* Compensate for simulation model behaviour */
3314 for (i = 0; i < RW_MGR_MEM_IF_READ_DQS_WIDTH; i++) {
3315 scc_mgr_set_dqs_bus_in_delay(i, 10);
3316 scc_mgr_load_dqs(i);
3318 writel(0, &sdr_scc_mgr->update);
3321 * ArriaV has hard FIFOs that can only be initialized by incrementing
3324 vfifo_offset = CALIB_VFIFO_OFFSET;
3325 for (j = 0; j < vfifo_offset; j++)
3326 writel(0xff, &phy_mgr_cmd->inc_vfifo_hard_phy);
3327 writel(0, &phy_mgr_cmd->fifo_reset);
3330 * For Arria V and Cyclone V with hard LFIFO, we get the skip-cal
3331 * setting from generation-time constant.
3333 gbl->curr_read_lat = CALIB_LFIFO_OFFSET;
3334 writel(gbl->curr_read_lat, &phy_mgr_cfg->phy_rlat);
3338 * mem_calibrate() - Memory calibration entry point.
3340 * Perform memory calibration.
3342 static uint32_t mem_calibrate(void)
3345 uint32_t rank_bgn, sr;
3346 uint32_t write_group, write_test_bgn;
3347 uint32_t read_group, read_test_bgn;
3348 uint32_t run_groups, current_run;
3349 uint32_t failing_groups = 0;
3350 uint32_t group_failed = 0;
3352 const u32 rwdqs_ratio = RW_MGR_MEM_IF_READ_DQS_WIDTH /
3353 RW_MGR_MEM_IF_WRITE_DQS_WIDTH;
3355 debug("%s:%d\n", __func__, __LINE__);
3357 /* Initialize the data settings */
3358 gbl->error_substage = CAL_SUBSTAGE_NIL;
3359 gbl->error_stage = CAL_STAGE_NIL;
3360 gbl->error_group = 0xff;
3364 /* Initialize WLAT and RLAT. */
3367 /* Initialize bit slips. */
3368 mem_precharge_and_activate();
3370 for (i = 0; i < RW_MGR_MEM_IF_READ_DQS_WIDTH; i++) {
3371 writel(i, SDR_PHYGRP_SCCGRP_ADDRESS |
3372 SCC_MGR_GROUP_COUNTER_OFFSET);
3373 /* Only needed once to set all groups, pins, DQ, DQS, DM. */
3375 scc_mgr_set_hhp_extras();
3377 scc_set_bypass_mode(i);
3380 /* Calibration is skipped. */
3381 if ((dyn_calib_steps & CALIB_SKIP_ALL) == CALIB_SKIP_ALL) {
3383 * Set VFIFO and LFIFO to instant-on settings in skip
3386 mem_skip_calibrate();
3389 * Do not remove this line as it makes sure all of our
3390 * decisions have been applied.
3392 writel(0, &sdr_scc_mgr->update);
3396 /* Calibration is not skipped. */
3397 for (i = 0; i < NUM_CALIB_REPEAT; i++) {
3399 * Zero all delay chain/phase settings for all
3400 * groups and all shadow register sets.
3404 run_groups = ~param->skip_groups;
3406 for (write_group = 0, write_test_bgn = 0; write_group
3407 < RW_MGR_MEM_IF_WRITE_DQS_WIDTH; write_group++,
3408 write_test_bgn += RW_MGR_MEM_DQ_PER_WRITE_DQS) {
3410 /* Initialize the group failure */
3413 current_run = run_groups & ((1 <<
3414 RW_MGR_NUM_DQS_PER_WRITE_GROUP) - 1);
3415 run_groups = run_groups >>
3416 RW_MGR_NUM_DQS_PER_WRITE_GROUP;
3418 if (current_run == 0)
3421 writel(write_group, SDR_PHYGRP_SCCGRP_ADDRESS |
3422 SCC_MGR_GROUP_COUNTER_OFFSET);
3423 scc_mgr_zero_group(write_group, 0);
3425 for (read_group = write_group * rwdqs_ratio,
3427 read_group < (write_group + 1) * rwdqs_ratio;
3429 read_test_bgn += RW_MGR_MEM_DQ_PER_READ_DQS) {
3430 if (STATIC_CALIB_STEPS & CALIB_SKIP_VFIFO)
3433 /* Calibrate the VFIFO */
3434 if (rw_mgr_mem_calibrate_vfifo(read_group,
3438 if (!(gbl->phy_debug_mode_flags & PHY_DEBUG_SWEEP_ALL_GROUPS))
3441 /* The group failed, we're done. */
3445 /* Calibrate the output side */
3446 for (rank_bgn = 0, sr = 0;
3447 rank_bgn < RW_MGR_MEM_NUMBER_OF_RANKS;
3448 rank_bgn += NUM_RANKS_PER_SHADOW_REG, sr++) {
3449 if (STATIC_CALIB_STEPS & CALIB_SKIP_WRITES)
3452 /* Not needed in quick mode! */
3453 if (STATIC_CALIB_STEPS & CALIB_SKIP_DELAY_SWEEPS)
3457 * Determine if this set of ranks
3458 * should be skipped entirely.
3460 if (param->skip_shadow_regs[sr])
3463 /* Calibrate WRITEs */
3464 if (rw_mgr_mem_calibrate_writes(rank_bgn,
3465 write_group, write_test_bgn))
3469 if (!(gbl->phy_debug_mode_flags & PHY_DEBUG_SWEEP_ALL_GROUPS))
3473 /* Some group failed, we're done. */
3477 for (read_group = write_group * rwdqs_ratio,
3479 read_group < (write_group + 1) * rwdqs_ratio;
3481 read_test_bgn += RW_MGR_MEM_DQ_PER_READ_DQS) {
3482 if (STATIC_CALIB_STEPS & CALIB_SKIP_WRITES)
3485 if (rw_mgr_mem_calibrate_vfifo_end(read_group,
3489 if (!(gbl->phy_debug_mode_flags & PHY_DEBUG_SWEEP_ALL_GROUPS))
3492 /* The group failed, we're done. */
3496 /* No group failed, continue as usual. */
3499 grp_failed: /* A group failed, increment the counter. */
3504 * USER If there are any failing groups then report
3507 if (failing_groups != 0)
3510 if (STATIC_CALIB_STEPS & CALIB_SKIP_LFIFO)
3514 * If we're skipping groups as part of debug,
3515 * don't calibrate LFIFO.
3517 if (param->skip_groups != 0)
3520 /* Calibrate the LFIFO */
3521 if (!rw_mgr_mem_calibrate_lfifo())
3526 * Do not remove this line as it makes sure all of our decisions
3527 * have been applied.
3529 writel(0, &sdr_scc_mgr->update);
3534 * run_mem_calibrate() - Perform memory calibration
3536 * This function triggers the entire memory calibration procedure.
3538 static int run_mem_calibrate(void)
3542 debug("%s:%d\n", __func__, __LINE__);
3544 /* Reset pass/fail status shown on afi_cal_success/fail */
3545 writel(PHY_MGR_CAL_RESET, &phy_mgr_cfg->cal_status);
3547 /* Stop tracking manager. */
3548 clrbits_le32(&sdr_ctrl->ctrl_cfg, 1 << 22);
3550 phy_mgr_initialize();
3551 rw_mgr_mem_initialize();
3553 /* Perform the actual memory calibration. */
3554 pass = mem_calibrate();
3556 mem_precharge_and_activate();
3557 writel(0, &phy_mgr_cmd->fifo_reset);
3560 rw_mgr_mem_handoff();
3562 * In Hard PHY this is a 2-bit control:
3564 * 1: DDIO Mux Select
3566 writel(0x2, &phy_mgr_cfg->mux_sel);
3568 /* Start tracking manager. */
3569 setbits_le32(&sdr_ctrl->ctrl_cfg, 1 << 22);
3575 * debug_mem_calibrate() - Report result of memory calibration
3576 * @pass: Value indicating whether calibration passed or failed
3578 * This function reports the results of the memory calibration
3579 * and writes debug information into the register file.
3581 static void debug_mem_calibrate(int pass)
3583 uint32_t debug_info;
3586 printf("%s: CALIBRATION PASSED\n", __FILE__);
3591 if (gbl->fom_in > 0xff)
3594 if (gbl->fom_out > 0xff)
3595 gbl->fom_out = 0xff;
3597 /* Update the FOM in the register file */
3598 debug_info = gbl->fom_in;
3599 debug_info |= gbl->fom_out << 8;
3600 writel(debug_info, &sdr_reg_file->fom);
3602 writel(debug_info, &phy_mgr_cfg->cal_debug_info);
3603 writel(PHY_MGR_CAL_SUCCESS, &phy_mgr_cfg->cal_status);
3605 printf("%s: CALIBRATION FAILED\n", __FILE__);
3607 debug_info = gbl->error_stage;
3608 debug_info |= gbl->error_substage << 8;
3609 debug_info |= gbl->error_group << 16;
3611 writel(debug_info, &sdr_reg_file->failing_stage);
3612 writel(debug_info, &phy_mgr_cfg->cal_debug_info);
3613 writel(PHY_MGR_CAL_FAIL, &phy_mgr_cfg->cal_status);
3615 /* Update the failing group/stage in the register file */
3616 debug_info = gbl->error_stage;
3617 debug_info |= gbl->error_substage << 8;
3618 debug_info |= gbl->error_group << 16;
3619 writel(debug_info, &sdr_reg_file->failing_stage);
3622 printf("%s: Calibration complete\n", __FILE__);
3626 * hc_initialize_rom_data() - Initialize ROM data
3628 * Initialize ROM data.
3630 static void hc_initialize_rom_data(void)
3634 addr = SDR_PHYGRP_RWMGRGRP_ADDRESS | RW_MGR_INST_ROM_WRITE_OFFSET;
3635 for (i = 0; i < ARRAY_SIZE(inst_rom_init); i++)
3636 writel(inst_rom_init[i], addr + (i << 2));
3638 addr = SDR_PHYGRP_RWMGRGRP_ADDRESS | RW_MGR_AC_ROM_WRITE_OFFSET;
3639 for (i = 0; i < ARRAY_SIZE(ac_rom_init); i++)
3640 writel(ac_rom_init[i], addr + (i << 2));
3644 * initialize_reg_file() - Initialize SDR register file
3646 * Initialize SDR register file.
3648 static void initialize_reg_file(void)
3650 /* Initialize the register file with the correct data */
3651 writel(REG_FILE_INIT_SEQ_SIGNATURE, &sdr_reg_file->signature);
3652 writel(0, &sdr_reg_file->debug_data_addr);
3653 writel(0, &sdr_reg_file->cur_stage);
3654 writel(0, &sdr_reg_file->fom);
3655 writel(0, &sdr_reg_file->failing_stage);
3656 writel(0, &sdr_reg_file->debug1);
3657 writel(0, &sdr_reg_file->debug2);
3661 * initialize_hps_phy() - Initialize HPS PHY
3663 * Initialize HPS PHY.
3665 static void initialize_hps_phy(void)
3669 * Tracking also gets configured here because it's in the
3672 uint32_t trk_sample_count = 7500;
3673 uint32_t trk_long_idle_sample_count = (10 << 16) | 100;
3675 * Format is number of outer loops in the 16 MSB, sample
3680 reg |= SDR_CTRLGRP_PHYCTRL_PHYCTRL_0_ACDELAYEN_SET(2);
3681 reg |= SDR_CTRLGRP_PHYCTRL_PHYCTRL_0_DQDELAYEN_SET(1);
3682 reg |= SDR_CTRLGRP_PHYCTRL_PHYCTRL_0_DQSDELAYEN_SET(1);
3683 reg |= SDR_CTRLGRP_PHYCTRL_PHYCTRL_0_DQSLOGICDELAYEN_SET(1);
3684 reg |= SDR_CTRLGRP_PHYCTRL_PHYCTRL_0_RESETDELAYEN_SET(0);
3685 reg |= SDR_CTRLGRP_PHYCTRL_PHYCTRL_0_LPDDRDIS_SET(1);
3687 * This field selects the intrinsic latency to RDATA_EN/FULL path.
3688 * 00-bypass, 01- add 5 cycles, 10- add 10 cycles, 11- add 15 cycles.
3690 reg |= SDR_CTRLGRP_PHYCTRL_PHYCTRL_0_ADDLATSEL_SET(0);
3691 reg |= SDR_CTRLGRP_PHYCTRL_PHYCTRL_0_SAMPLECOUNT_19_0_SET(
3693 writel(reg, &sdr_ctrl->phy_ctrl0);
3696 reg |= SDR_CTRLGRP_PHYCTRL_PHYCTRL_1_SAMPLECOUNT_31_20_SET(
3698 SDR_CTRLGRP_PHYCTRL_PHYCTRL_0_SAMPLECOUNT_19_0_WIDTH);
3699 reg |= SDR_CTRLGRP_PHYCTRL_PHYCTRL_1_LONGIDLESAMPLECOUNT_19_0_SET(
3700 trk_long_idle_sample_count);
3701 writel(reg, &sdr_ctrl->phy_ctrl1);
3704 reg |= SDR_CTRLGRP_PHYCTRL_PHYCTRL_2_LONGIDLESAMPLECOUNT_31_20_SET(
3705 trk_long_idle_sample_count >>
3706 SDR_CTRLGRP_PHYCTRL_PHYCTRL_1_LONGIDLESAMPLECOUNT_19_0_WIDTH);
3707 writel(reg, &sdr_ctrl->phy_ctrl2);
3711 * initialize_tracking() - Initialize tracking
3713 * Initialize the register file with usable initial data.
3715 static void initialize_tracking(void)
3718 * Initialize the register file with the correct data.
3719 * Compute usable version of value in case we skip full
3720 * computation later.
3722 writel(DIV_ROUND_UP(IO_DELAY_PER_OPA_TAP, IO_DELAY_PER_DCHAIN_TAP) - 1,
3723 &sdr_reg_file->dtaps_per_ptap);
3725 /* trk_sample_count */
3726 writel(7500, &sdr_reg_file->trk_sample_count);
3728 /* longidle outer loop [15:0] */
3729 writel((10 << 16) | (100 << 0), &sdr_reg_file->trk_longidle);
3732 * longidle sample count [31:24]
3733 * trfc, worst case of 933Mhz 4Gb [23:16]
3734 * trcd, worst case [15:8]
3737 writel((243 << 24) | (14 << 16) | (10 << 8) | (4 << 0),
3738 &sdr_reg_file->delays);
3741 writel((RW_MGR_IDLE << 24) | (RW_MGR_ACTIVATE_1 << 16) |
3742 (RW_MGR_SGLE_READ << 8) | (RW_MGR_PRECHARGE_ALL << 0),
3743 &sdr_reg_file->trk_rw_mgr_addr);
3745 writel(RW_MGR_MEM_IF_READ_DQS_WIDTH,
3746 &sdr_reg_file->trk_read_dqs_width);
3749 writel((RW_MGR_REFRESH_ALL << 24) | (1000 << 0),
3750 &sdr_reg_file->trk_rfsh);
3753 int sdram_calibration_full(void)
3755 struct param_type my_param;
3756 struct gbl_type my_gbl;
3759 memset(&my_param, 0, sizeof(my_param));
3760 memset(&my_gbl, 0, sizeof(my_gbl));
3765 /* Set the calibration enabled by default */
3766 gbl->phy_debug_mode_flags |= PHY_DEBUG_ENABLE_CAL_RPT;
3768 * Only sweep all groups (regardless of fail state) by default
3769 * Set enabled read test by default.
3771 #if DISABLE_GUARANTEED_READ
3772 gbl->phy_debug_mode_flags |= PHY_DEBUG_DISABLE_GUARANTEED_READ;
3774 /* Initialize the register file */
3775 initialize_reg_file();
3777 /* Initialize any PHY CSR */
3778 initialize_hps_phy();
3780 scc_mgr_initialize();
3782 initialize_tracking();
3784 printf("%s: Preparing to start memory calibration\n", __FILE__);
3786 debug("%s:%d\n", __func__, __LINE__);
3787 debug_cond(DLEVEL == 1,
3788 "DDR3 FULL_RATE ranks=%u cs/dimm=%u dq/dqs=%u,%u vg/dqs=%u,%u ",
3789 RW_MGR_MEM_NUMBER_OF_RANKS, RW_MGR_MEM_NUMBER_OF_CS_PER_DIMM,
3790 RW_MGR_MEM_DQ_PER_READ_DQS, RW_MGR_MEM_DQ_PER_WRITE_DQS,
3791 RW_MGR_MEM_VIRTUAL_GROUPS_PER_READ_DQS,
3792 RW_MGR_MEM_VIRTUAL_GROUPS_PER_WRITE_DQS);
3793 debug_cond(DLEVEL == 1,
3794 "dqs=%u,%u dq=%u dm=%u ptap_delay=%u dtap_delay=%u ",
3795 RW_MGR_MEM_IF_READ_DQS_WIDTH, RW_MGR_MEM_IF_WRITE_DQS_WIDTH,
3796 RW_MGR_MEM_DATA_WIDTH, RW_MGR_MEM_DATA_MASK_WIDTH,
3797 IO_DELAY_PER_OPA_TAP, IO_DELAY_PER_DCHAIN_TAP);
3798 debug_cond(DLEVEL == 1, "dtap_dqsen_delay=%u, dll=%u",
3799 IO_DELAY_PER_DQS_EN_DCHAIN_TAP, IO_DLL_CHAIN_LENGTH);
3800 debug_cond(DLEVEL == 1, "max values: en_p=%u dqdqs_p=%u en_d=%u dqs_in_d=%u ",
3801 IO_DQS_EN_PHASE_MAX, IO_DQDQS_OUT_PHASE_MAX,
3802 IO_DQS_EN_DELAY_MAX, IO_DQS_IN_DELAY_MAX);
3803 debug_cond(DLEVEL == 1, "io_in_d=%u io_out1_d=%u io_out2_d=%u ",
3804 IO_IO_IN_DELAY_MAX, IO_IO_OUT1_DELAY_MAX,
3805 IO_IO_OUT2_DELAY_MAX);
3806 debug_cond(DLEVEL == 1, "dqs_in_reserve=%u dqs_out_reserve=%u\n",
3807 IO_DQS_IN_RESERVE, IO_DQS_OUT_RESERVE);
3809 hc_initialize_rom_data();
3811 /* update info for sims */
3812 reg_file_set_stage(CAL_STAGE_NIL);
3813 reg_file_set_group(0);
3816 * Load global needed for those actions that require
3817 * some dynamic calibration support.
3819 dyn_calib_steps = STATIC_CALIB_STEPS;
3821 * Load global to allow dynamic selection of delay loop settings
3822 * based on calibration mode.
3824 if (!(dyn_calib_steps & CALIB_SKIP_DELAY_LOOPS))
3825 skip_delay_mask = 0xff;
3827 skip_delay_mask = 0x0;
3829 pass = run_mem_calibrate();
3830 debug_mem_calibrate(pass);