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.
1177 static uint32_t rw_mgr_mem_calibrate_read_test(uint32_t rank_bgn, uint32_t group,
1178 uint32_t num_tries, uint32_t all_correct, uint32_t *bit_chk,
1179 uint32_t all_groups, uint32_t all_ranks)
1182 uint32_t correct_mask_vg;
1183 uint32_t tmp_bit_chk;
1184 uint32_t rank_end = all_ranks ? RW_MGR_MEM_NUMBER_OF_RANKS :
1185 (rank_bgn + NUM_RANKS_PER_SHADOW_REG);
1187 uint32_t base_rw_mgr;
1189 *bit_chk = param->read_correct_mask;
1190 correct_mask_vg = param->read_correct_mask_vg;
1192 uint32_t quick_read_mode = (((STATIC_CALIB_STEPS) &
1193 CALIB_SKIP_DELAY_SWEEPS) && ENABLE_SUPER_QUICK_CALIBRATION);
1195 for (r = rank_bgn; r < rank_end; r++) {
1196 if (param->skip_ranks[r])
1197 /* request to skip the rank */
1201 set_rank_and_odt_mask(r, RW_MGR_ODT_MODE_READ_WRITE);
1203 writel(0x10, &sdr_rw_load_mgr_regs->load_cntr1);
1205 writel(RW_MGR_READ_B2B_WAIT1,
1206 &sdr_rw_load_jump_mgr_regs->load_jump_add1);
1208 writel(0x10, &sdr_rw_load_mgr_regs->load_cntr2);
1209 writel(RW_MGR_READ_B2B_WAIT2,
1210 &sdr_rw_load_jump_mgr_regs->load_jump_add2);
1212 if (quick_read_mode)
1213 writel(0x1, &sdr_rw_load_mgr_regs->load_cntr0);
1214 /* need at least two (1+1) reads to capture failures */
1215 else if (all_groups)
1216 writel(0x06, &sdr_rw_load_mgr_regs->load_cntr0);
1218 writel(0x32, &sdr_rw_load_mgr_regs->load_cntr0);
1220 writel(RW_MGR_READ_B2B,
1221 &sdr_rw_load_jump_mgr_regs->load_jump_add0);
1223 writel(RW_MGR_MEM_IF_READ_DQS_WIDTH *
1224 RW_MGR_MEM_VIRTUAL_GROUPS_PER_READ_DQS - 1,
1225 &sdr_rw_load_mgr_regs->load_cntr3);
1227 writel(0x0, &sdr_rw_load_mgr_regs->load_cntr3);
1229 writel(RW_MGR_READ_B2B,
1230 &sdr_rw_load_jump_mgr_regs->load_jump_add3);
1233 for (vg = RW_MGR_MEM_VIRTUAL_GROUPS_PER_READ_DQS-1; ; vg--) {
1234 /* reset the fifos to get pointers to known state */
1235 writel(0, &phy_mgr_cmd->fifo_reset);
1236 writel(0, SDR_PHYGRP_RWMGRGRP_ADDRESS |
1237 RW_MGR_RESET_READ_DATAPATH_OFFSET);
1239 tmp_bit_chk = tmp_bit_chk << (RW_MGR_MEM_DQ_PER_READ_DQS
1240 / RW_MGR_MEM_VIRTUAL_GROUPS_PER_READ_DQS);
1243 addr = SDR_PHYGRP_RWMGRGRP_ADDRESS | RW_MGR_RUN_ALL_GROUPS_OFFSET;
1245 addr = SDR_PHYGRP_RWMGRGRP_ADDRESS | RW_MGR_RUN_SINGLE_GROUP_OFFSET;
1247 writel(RW_MGR_READ_B2B, addr +
1248 ((group * RW_MGR_MEM_VIRTUAL_GROUPS_PER_READ_DQS +
1251 base_rw_mgr = readl(SDR_PHYGRP_RWMGRGRP_ADDRESS);
1252 tmp_bit_chk = tmp_bit_chk | (correct_mask_vg & ~(base_rw_mgr));
1257 *bit_chk &= tmp_bit_chk;
1260 addr = SDR_PHYGRP_RWMGRGRP_ADDRESS | RW_MGR_RUN_SINGLE_GROUP_OFFSET;
1261 writel(RW_MGR_CLEAR_DQS_ENABLE, addr + (group << 2));
1264 set_rank_and_odt_mask(0, RW_MGR_ODT_MODE_OFF);
1265 debug_cond(DLEVEL == 2, "%s:%d read_test(%u,ALL,%u) =>\
1266 (%u == %u) => %lu", __func__, __LINE__, group,
1267 all_groups, *bit_chk, param->read_correct_mask,
1268 (long unsigned int)(*bit_chk ==
1269 param->read_correct_mask));
1270 return *bit_chk == param->read_correct_mask;
1272 set_rank_and_odt_mask(0, RW_MGR_ODT_MODE_OFF);
1273 debug_cond(DLEVEL == 2, "%s:%d read_test(%u,ONE,%u) =>\
1274 (%u != %lu) => %lu\n", __func__, __LINE__,
1275 group, all_groups, *bit_chk, (long unsigned int)0,
1276 (long unsigned int)(*bit_chk != 0x00));
1277 return *bit_chk != 0x00;
1281 static uint32_t rw_mgr_mem_calibrate_read_test_all_ranks(uint32_t group,
1282 uint32_t num_tries, uint32_t all_correct, uint32_t *bit_chk,
1283 uint32_t all_groups)
1285 return rw_mgr_mem_calibrate_read_test(0, group, num_tries, all_correct,
1286 bit_chk, all_groups, 1);
1289 static void rw_mgr_incr_vfifo(uint32_t grp, uint32_t *v)
1291 writel(grp, &phy_mgr_cmd->inc_vfifo_hard_phy);
1295 static void rw_mgr_decr_vfifo(uint32_t grp, uint32_t *v)
1299 for (i = 0; i < VFIFO_SIZE-1; i++)
1300 rw_mgr_incr_vfifo(grp, v);
1303 static int find_vfifo_read(uint32_t grp, uint32_t *bit_chk)
1306 uint32_t fail_cnt = 0;
1307 uint32_t test_status;
1309 for (v = 0; v < VFIFO_SIZE; ) {
1310 debug_cond(DLEVEL == 2, "%s:%d find_dqs_en_phase: vfifo %u\n",
1311 __func__, __LINE__, v);
1312 test_status = rw_mgr_mem_calibrate_read_test_all_ranks
1313 (grp, 1, PASS_ONE_BIT, bit_chk, 0);
1321 /* fiddle with FIFO */
1322 rw_mgr_incr_vfifo(grp, &v);
1325 if (v >= VFIFO_SIZE) {
1326 /* no failing read found!! Something must have gone wrong */
1327 debug_cond(DLEVEL == 2, "%s:%d find_dqs_en_phase: vfifo failed\n",
1328 __func__, __LINE__);
1336 * sdr_find_phase() - Find DQS enable phase
1337 * @working: If 1, look for working phase, if 0, look for non-working phase
1338 * @grp: Read/Write group
1340 * @work: Working window position
1342 * @p: DQS Phase Iterator
1344 * Find working or non-working DQS enable phase setting.
1346 static int sdr_find_phase(int working, const u32 grp, u32 *v, u32 *work,
1350 const u32 end = VFIFO_SIZE + (working ? 0 : 1);
1352 for (; *i < end; (*i)++) {
1356 for (; *p <= IO_DQS_EN_PHASE_MAX; (*p)++) {
1357 scc_mgr_set_dqs_en_phase_all_ranks(grp, *p);
1359 ret = rw_mgr_mem_calibrate_read_test_all_ranks(grp, 1,
1360 PASS_ONE_BIT, &bit_chk, 0);
1367 *work += IO_DELAY_PER_OPA_TAP;
1370 if (*p > IO_DQS_EN_PHASE_MAX) {
1371 /* Fiddle with FIFO. */
1372 rw_mgr_incr_vfifo(grp, v);
1381 static int sdr_working_phase(uint32_t grp, uint32_t *work_bgn,
1382 uint32_t *v, uint32_t *d, uint32_t *p,
1385 const u32 dtaps_per_ptap = IO_DELAY_PER_OPA_TAP /
1386 IO_DELAY_PER_DQS_EN_DCHAIN_TAP;
1391 for (*d = 0; *d <= dtaps_per_ptap; (*d)++) {
1393 scc_mgr_set_dqs_en_delay_all_ranks(grp, *d);
1394 ret = sdr_find_phase(1, grp, v, work_bgn, i, p);
1397 *work_bgn += IO_DELAY_PER_DQS_EN_DCHAIN_TAP;
1400 /* Cannot find working solution */
1401 debug_cond(DLEVEL == 2, "%s:%d find_dqs_en_phase: no vfifo/ptap/dtap\n",
1402 __func__, __LINE__);
1406 static void sdr_backup_phase(uint32_t grp,
1407 uint32_t *work_bgn, uint32_t *v, uint32_t *d,
1413 /* Special case code for backing up a phase */
1415 *p = IO_DQS_EN_PHASE_MAX;
1416 rw_mgr_decr_vfifo(grp, v);
1420 tmp_delay = *work_bgn - IO_DELAY_PER_OPA_TAP;
1421 scc_mgr_set_dqs_en_phase_all_ranks(grp, *p);
1423 for (*d = 0; *d <= IO_DQS_EN_DELAY_MAX && tmp_delay < *work_bgn;
1424 (*d)++, tmp_delay += IO_DELAY_PER_DQS_EN_DCHAIN_TAP) {
1425 scc_mgr_set_dqs_en_delay_all_ranks(grp, *d);
1427 if (rw_mgr_mem_calibrate_read_test_all_ranks(grp, 1,
1430 *work_bgn = tmp_delay;
1436 * Restore VFIFO to old state before we decremented it
1440 if (*p > IO_DQS_EN_PHASE_MAX) {
1442 rw_mgr_incr_vfifo(grp, v);
1445 scc_mgr_set_dqs_en_delay_all_ranks(grp, 0);
1448 static int sdr_nonworking_phase(uint32_t grp, uint32_t *v, uint32_t *d,
1449 uint32_t *p, uint32_t *i, uint32_t *work_end)
1454 *work_end += IO_DELAY_PER_OPA_TAP;
1455 if (*p > IO_DQS_EN_PHASE_MAX) {
1456 /* Fiddle with FIFO. */
1458 rw_mgr_incr_vfifo(grp, v);
1461 ret = sdr_find_phase(0, grp, v, work_end, i, p);
1463 /* Cannot see edge of failing read. */
1464 debug_cond(DLEVEL == 2, "%s:%d: end: failed\n",
1465 __func__, __LINE__);
1472 * sdr_find_window_center() - Find center of the working DQS window.
1473 * @grp: Read/Write group
1474 * @work_bgn: First working settings
1475 * @work_end: Last working settings
1478 * Find center of the working DQS enable window.
1480 static int sdr_find_window_center(const u32 grp, const u32 work_bgn,
1481 const u32 work_end, const u32 val)
1483 u32 bit_chk, work_mid, v = val;
1487 work_mid = (work_bgn + work_end) / 2;
1489 debug_cond(DLEVEL == 2, "work_bgn=%d work_end=%d work_mid=%d\n",
1490 work_bgn, work_end, work_mid);
1491 /* Get the middle delay to be less than a VFIFO delay */
1492 tmp_delay = (IO_DQS_EN_PHASE_MAX + 1) * IO_DELAY_PER_OPA_TAP;
1494 debug_cond(DLEVEL == 2, "vfifo ptap delay %d\n", tmp_delay);
1495 work_mid %= tmp_delay;
1496 debug_cond(DLEVEL == 2, "new work_mid %d\n", work_mid);
1498 tmp_delay = rounddown(work_mid, IO_DELAY_PER_OPA_TAP);
1499 if (tmp_delay > IO_DQS_EN_PHASE_MAX * IO_DELAY_PER_OPA_TAP)
1500 tmp_delay = IO_DQS_EN_PHASE_MAX * IO_DELAY_PER_OPA_TAP;
1501 p = tmp_delay / IO_DELAY_PER_OPA_TAP;
1503 debug_cond(DLEVEL == 2, "new p %d, tmp_delay=%d\n", p, tmp_delay);
1505 d = DIV_ROUND_UP(work_mid - tmp_delay, IO_DELAY_PER_DQS_EN_DCHAIN_TAP);
1506 if (d > IO_DQS_EN_DELAY_MAX)
1507 d = IO_DQS_EN_DELAY_MAX;
1508 tmp_delay += d * IO_DELAY_PER_DQS_EN_DCHAIN_TAP;
1510 debug_cond(DLEVEL == 2, "new d %d, tmp_delay=%d\n", d, tmp_delay);
1512 scc_mgr_set_dqs_en_phase_all_ranks(grp, p);
1513 scc_mgr_set_dqs_en_delay_all_ranks(grp, d);
1516 * push vfifo until we can successfully calibrate. We can do this
1517 * because the largest possible margin in 1 VFIFO cycle.
1519 for (i = 0; i < VFIFO_SIZE; i++) {
1520 debug_cond(DLEVEL == 2, "find_dqs_en_phase: center: vfifo=%u\n",
1522 if (rw_mgr_mem_calibrate_read_test_all_ranks(grp, 1,
1525 debug_cond(DLEVEL == 2,
1526 "%s:%d center: found: vfifo=%u ptap=%u dtap=%u\n",
1527 __func__, __LINE__, v, p, d);
1531 /* Fiddle with FIFO. */
1532 rw_mgr_incr_vfifo(grp, &v);
1535 debug_cond(DLEVEL == 2, "%s:%d center: failed.\n",
1536 __func__, __LINE__);
1540 /* find a good dqs enable to use */
1541 static uint32_t rw_mgr_mem_calibrate_vfifo_find_dqs_en_phase(uint32_t grp)
1543 uint32_t v, d, p, i;
1545 uint32_t dtaps_per_ptap;
1546 uint32_t work_bgn, work_end;
1547 uint32_t found_passing_read, found_failing_read, initial_failing_dtap;
1549 debug("%s:%d %u\n", __func__, __LINE__, grp);
1551 reg_file_set_sub_stage(CAL_SUBSTAGE_VFIFO_CENTER);
1553 scc_mgr_set_dqs_en_delay_all_ranks(grp, 0);
1554 scc_mgr_set_dqs_en_phase_all_ranks(grp, 0);
1556 /* ************************************************************** */
1557 /* * Step 0 : Determine number of delay taps for each phase tap * */
1558 dtaps_per_ptap = IO_DELAY_PER_OPA_TAP/IO_DELAY_PER_DQS_EN_DCHAIN_TAP;
1560 /* ********************************************************* */
1561 /* * Step 1 : First push vfifo until we get a failing read * */
1562 v = find_vfifo_read(grp, &bit_chk);
1564 /* ******************************************************** */
1565 /* * step 2: find first working phase, increment in ptaps * */
1567 if (sdr_working_phase(grp, &work_bgn, &v, &d, &p, &i))
1570 work_end = work_bgn;
1573 * If d is 0 then the working window covers a phase tap and
1574 * we can follow the old procedure otherwise, we've found the beginning,
1575 * and we need to increment the dtaps until we find the end.
1578 /* ********************************************************* */
1579 /* * step 3a: if we have room, back off by one and
1580 increment in dtaps * */
1582 sdr_backup_phase(grp, &work_bgn, &v, &d, &p);
1584 /* ********************************************************* */
1585 /* * step 4a: go forward from working phase to non working
1586 phase, increment in ptaps * */
1587 if (sdr_nonworking_phase(grp, &v, &d, &p, &i, &work_end))
1590 /* ********************************************************* */
1591 /* * step 5a: back off one from last, increment in dtaps * */
1593 /* Special case code for backing up a phase */
1595 p = IO_DQS_EN_PHASE_MAX;
1596 rw_mgr_decr_vfifo(grp, &v);
1601 work_end -= IO_DELAY_PER_OPA_TAP;
1602 scc_mgr_set_dqs_en_phase_all_ranks(grp, p);
1604 /* * The actual increment of dtaps is done outside of
1605 the if/else loop to share code */
1608 debug_cond(DLEVEL == 2, "%s:%d find_dqs_en_phase: v/p: \
1609 vfifo=%u ptap=%u\n", __func__, __LINE__,
1612 /* ******************************************************* */
1613 /* * step 3-5b: Find the right edge of the window using
1615 debug_cond(DLEVEL == 2, "%s:%d find_dqs_en_phase:vfifo=%u \
1616 ptap=%u dtap=%u bgn=%u\n", __func__, __LINE__,
1619 work_end = work_bgn;
1622 /* The dtap increment to find the failing edge is done here */
1623 for (; d <= IO_DQS_EN_DELAY_MAX; d++, work_end +=
1624 IO_DELAY_PER_DQS_EN_DCHAIN_TAP) {
1625 debug_cond(DLEVEL == 2, "%s:%d find_dqs_en_phase: \
1626 end-2: dtap=%u\n", __func__, __LINE__, d);
1627 scc_mgr_set_dqs_en_delay_all_ranks(grp, d);
1629 if (!rw_mgr_mem_calibrate_read_test_all_ranks(grp, 1,
1636 /* Go back to working dtap */
1638 work_end -= IO_DELAY_PER_DQS_EN_DCHAIN_TAP;
1640 debug_cond(DLEVEL == 2, "%s:%d find_dqs_en_phase: v/p/d: vfifo=%u \
1641 ptap=%u dtap=%u end=%u\n", __func__, __LINE__,
1642 v, p, d-1, work_end);
1644 if (work_end < work_bgn) {
1646 debug_cond(DLEVEL == 2, "%s:%d find_dqs_en_phase: end-2: \
1647 failed\n", __func__, __LINE__);
1651 debug_cond(DLEVEL == 2, "%s:%d find_dqs_en_phase: found range [%u,%u]\n",
1652 __func__, __LINE__, work_bgn, work_end);
1654 /* *************************************************************** */
1656 * * We need to calculate the number of dtaps that equal a ptap
1657 * * To do that we'll back up a ptap and re-find the edge of the
1658 * * window using dtaps
1661 debug_cond(DLEVEL == 2, "%s:%d find_dqs_en_phase: calculate dtaps_per_ptap \
1662 for tracking\n", __func__, __LINE__);
1664 /* Special case code for backing up a phase */
1666 p = IO_DQS_EN_PHASE_MAX;
1667 rw_mgr_decr_vfifo(grp, &v);
1668 debug_cond(DLEVEL == 2, "%s:%d find_dqs_en_phase: backedup \
1669 cycle/phase: v=%u p=%u\n", __func__, __LINE__,
1673 debug_cond(DLEVEL == 2, "%s:%d find_dqs_en_phase: backedup \
1674 phase only: v=%u p=%u", __func__, __LINE__,
1678 scc_mgr_set_dqs_en_phase_all_ranks(grp, p);
1681 * Increase dtap until we first see a passing read (in case the
1682 * window is smaller than a ptap),
1683 * and then a failing read to mark the edge of the window again
1686 /* Find a passing read */
1687 debug_cond(DLEVEL == 2, "%s:%d find_dqs_en_phase: find passing read\n",
1688 __func__, __LINE__);
1689 found_passing_read = 0;
1690 found_failing_read = 0;
1691 initial_failing_dtap = d;
1692 for (; d <= IO_DQS_EN_DELAY_MAX; d++) {
1693 debug_cond(DLEVEL == 2, "%s:%d find_dqs_en_phase: testing \
1694 read d=%u\n", __func__, __LINE__, d);
1695 scc_mgr_set_dqs_en_delay_all_ranks(grp, d);
1697 if (rw_mgr_mem_calibrate_read_test_all_ranks(grp, 1,
1700 found_passing_read = 1;
1705 if (found_passing_read) {
1706 /* Find a failing read */
1707 debug_cond(DLEVEL == 2, "%s:%d find_dqs_en_phase: find failing \
1708 read\n", __func__, __LINE__);
1709 for (d = d + 1; d <= IO_DQS_EN_DELAY_MAX; d++) {
1710 debug_cond(DLEVEL == 2, "%s:%d find_dqs_en_phase: \
1711 testing read d=%u\n", __func__, __LINE__, d);
1712 scc_mgr_set_dqs_en_delay_all_ranks(grp, d);
1714 if (!rw_mgr_mem_calibrate_read_test_all_ranks
1715 (grp, 1, PASS_ONE_BIT, &bit_chk, 0)) {
1716 found_failing_read = 1;
1721 debug_cond(DLEVEL == 1, "%s:%d find_dqs_en_phase: failed to \
1722 calculate dtaps", __func__, __LINE__);
1723 debug_cond(DLEVEL == 1, "per ptap. Fall back on static value\n");
1727 * The dynamically calculated dtaps_per_ptap is only valid if we
1728 * found a passing/failing read. If we didn't, it means d hit the max
1729 * (IO_DQS_EN_DELAY_MAX). Otherwise, dtaps_per_ptap retains its
1730 * statically calculated value.
1732 if (found_passing_read && found_failing_read)
1733 dtaps_per_ptap = d - initial_failing_dtap;
1735 writel(dtaps_per_ptap, &sdr_reg_file->dtaps_per_ptap);
1736 debug_cond(DLEVEL == 2, "%s:%d find_dqs_en_phase: dtaps_per_ptap=%u \
1737 - %u = %u", __func__, __LINE__, d,
1738 initial_failing_dtap, dtaps_per_ptap);
1740 /* ******************************************** */
1741 /* * step 6: Find the centre of the window * */
1742 if (sdr_find_window_centre(grp, work_bgn, work_end, v))
1743 return 0; /* FIXME: Old code, return 0 means failure :-( */
1748 /* per-bit deskew DQ and center */
1749 static uint32_t rw_mgr_mem_calibrate_vfifo_center(uint32_t rank_bgn,
1750 uint32_t write_group, uint32_t read_group, uint32_t test_bgn,
1751 uint32_t use_read_test, uint32_t update_fom)
1753 uint32_t i, p, d, min_index;
1755 * Store these as signed since there are comparisons with
1759 uint32_t sticky_bit_chk;
1760 int32_t left_edge[RW_MGR_MEM_DQ_PER_READ_DQS];
1761 int32_t right_edge[RW_MGR_MEM_DQ_PER_READ_DQS];
1762 int32_t final_dq[RW_MGR_MEM_DQ_PER_READ_DQS];
1764 int32_t orig_mid_min, mid_min;
1765 int32_t new_dqs, start_dqs, start_dqs_en, shift_dq, final_dqs,
1767 int32_t dq_margin, dqs_margin;
1769 uint32_t temp_dq_in_delay1, temp_dq_in_delay2;
1772 debug("%s:%d: %u %u", __func__, __LINE__, read_group, test_bgn);
1774 addr = SDR_PHYGRP_SCCGRP_ADDRESS | SCC_MGR_DQS_IN_DELAY_OFFSET;
1775 start_dqs = readl(addr + (read_group << 2));
1776 if (IO_SHIFT_DQS_EN_WHEN_SHIFT_DQS)
1777 start_dqs_en = readl(addr + ((read_group << 2)
1778 - IO_DQS_EN_DELAY_OFFSET));
1780 /* set the left and right edge of each bit to an illegal value */
1781 /* use (IO_IO_IN_DELAY_MAX + 1) as an illegal value */
1783 for (i = 0; i < RW_MGR_MEM_DQ_PER_READ_DQS; i++) {
1784 left_edge[i] = IO_IO_IN_DELAY_MAX + 1;
1785 right_edge[i] = IO_IO_IN_DELAY_MAX + 1;
1788 /* Search for the left edge of the window for each bit */
1789 for (d = 0; d <= IO_IO_IN_DELAY_MAX; d++) {
1790 scc_mgr_apply_group_dq_in_delay(write_group, test_bgn, d);
1792 writel(0, &sdr_scc_mgr->update);
1795 * Stop searching when the read test doesn't pass AND when
1796 * we've seen a passing read on every bit.
1798 if (use_read_test) {
1799 stop = !rw_mgr_mem_calibrate_read_test(rank_bgn,
1800 read_group, NUM_READ_PB_TESTS, PASS_ONE_BIT,
1803 rw_mgr_mem_calibrate_write_test(rank_bgn, write_group,
1806 bit_chk = bit_chk >> (RW_MGR_MEM_DQ_PER_READ_DQS *
1807 (read_group - (write_group *
1808 RW_MGR_MEM_IF_READ_DQS_WIDTH /
1809 RW_MGR_MEM_IF_WRITE_DQS_WIDTH)));
1810 stop = (bit_chk == 0);
1812 sticky_bit_chk = sticky_bit_chk | bit_chk;
1813 stop = stop && (sticky_bit_chk == param->read_correct_mask);
1814 debug_cond(DLEVEL == 2, "%s:%d vfifo_center(left): dtap=%u => %u == %u \
1815 && %u", __func__, __LINE__, d,
1817 param->read_correct_mask, stop);
1822 for (i = 0; i < RW_MGR_MEM_DQ_PER_READ_DQS; i++) {
1824 /* Remember a passing test as the
1828 /* If a left edge has not been seen yet,
1829 then a future passing test will mark
1830 this edge as the right edge */
1832 IO_IO_IN_DELAY_MAX + 1) {
1833 right_edge[i] = -(d + 1);
1836 bit_chk = bit_chk >> 1;
1841 /* Reset DQ delay chains to 0 */
1842 scc_mgr_apply_group_dq_in_delay(test_bgn, 0);
1844 for (i = RW_MGR_MEM_DQ_PER_READ_DQS - 1;; i--) {
1845 debug_cond(DLEVEL == 2, "%s:%d vfifo_center: left_edge[%u]: \
1846 %d right_edge[%u]: %d\n", __func__, __LINE__,
1847 i, left_edge[i], i, right_edge[i]);
1850 * Check for cases where we haven't found the left edge,
1851 * which makes our assignment of the the right edge invalid.
1852 * Reset it to the illegal value.
1854 if ((left_edge[i] == IO_IO_IN_DELAY_MAX + 1) && (
1855 right_edge[i] != IO_IO_IN_DELAY_MAX + 1)) {
1856 right_edge[i] = IO_IO_IN_DELAY_MAX + 1;
1857 debug_cond(DLEVEL == 2, "%s:%d vfifo_center: reset \
1858 right_edge[%u]: %d\n", __func__, __LINE__,
1863 * Reset sticky bit (except for bits where we have seen
1864 * both the left and right edge).
1866 sticky_bit_chk = sticky_bit_chk << 1;
1867 if ((left_edge[i] != IO_IO_IN_DELAY_MAX + 1) &&
1868 (right_edge[i] != IO_IO_IN_DELAY_MAX + 1)) {
1869 sticky_bit_chk = sticky_bit_chk | 1;
1876 /* Search for the right edge of the window for each bit */
1877 for (d = 0; d <= IO_DQS_IN_DELAY_MAX - start_dqs; d++) {
1878 scc_mgr_set_dqs_bus_in_delay(read_group, d + start_dqs);
1879 if (IO_SHIFT_DQS_EN_WHEN_SHIFT_DQS) {
1880 uint32_t delay = d + start_dqs_en;
1881 if (delay > IO_DQS_EN_DELAY_MAX)
1882 delay = IO_DQS_EN_DELAY_MAX;
1883 scc_mgr_set_dqs_en_delay(read_group, delay);
1885 scc_mgr_load_dqs(read_group);
1887 writel(0, &sdr_scc_mgr->update);
1890 * Stop searching when the read test doesn't pass AND when
1891 * we've seen a passing read on every bit.
1893 if (use_read_test) {
1894 stop = !rw_mgr_mem_calibrate_read_test(rank_bgn,
1895 read_group, NUM_READ_PB_TESTS, PASS_ONE_BIT,
1898 rw_mgr_mem_calibrate_write_test(rank_bgn, write_group,
1901 bit_chk = bit_chk >> (RW_MGR_MEM_DQ_PER_READ_DQS *
1902 (read_group - (write_group *
1903 RW_MGR_MEM_IF_READ_DQS_WIDTH /
1904 RW_MGR_MEM_IF_WRITE_DQS_WIDTH)));
1905 stop = (bit_chk == 0);
1907 sticky_bit_chk = sticky_bit_chk | bit_chk;
1908 stop = stop && (sticky_bit_chk == param->read_correct_mask);
1910 debug_cond(DLEVEL == 2, "%s:%d vfifo_center(right): dtap=%u => %u == \
1911 %u && %u", __func__, __LINE__, d,
1912 sticky_bit_chk, param->read_correct_mask, stop);
1917 for (i = 0; i < RW_MGR_MEM_DQ_PER_READ_DQS; i++) {
1919 /* Remember a passing test as
1924 /* If a right edge has not been
1925 seen yet, then a future passing
1926 test will mark this edge as the
1928 if (right_edge[i] ==
1929 IO_IO_IN_DELAY_MAX + 1) {
1930 left_edge[i] = -(d + 1);
1933 /* d = 0 failed, but it passed
1934 when testing the left edge,
1935 so it must be marginal,
1937 if (right_edge[i] ==
1938 IO_IO_IN_DELAY_MAX + 1 &&
1944 /* If a right edge has not been
1945 seen yet, then a future passing
1946 test will mark this edge as the
1948 else if (right_edge[i] ==
1949 IO_IO_IN_DELAY_MAX +
1951 left_edge[i] = -(d + 1);
1956 debug_cond(DLEVEL == 2, "%s:%d vfifo_center[r,\
1957 d=%u]: ", __func__, __LINE__, d);
1958 debug_cond(DLEVEL == 2, "bit_chk_test=%d left_edge[%u]: %d ",
1959 (int)(bit_chk & 1), i, left_edge[i]);
1960 debug_cond(DLEVEL == 2, "right_edge[%u]: %d\n", i,
1962 bit_chk = bit_chk >> 1;
1967 /* Check that all bits have a window */
1968 for (i = 0; i < RW_MGR_MEM_DQ_PER_READ_DQS; i++) {
1969 debug_cond(DLEVEL == 2, "%s:%d vfifo_center: left_edge[%u]: \
1970 %d right_edge[%u]: %d", __func__, __LINE__,
1971 i, left_edge[i], i, right_edge[i]);
1972 if ((left_edge[i] == IO_IO_IN_DELAY_MAX + 1) || (right_edge[i]
1973 == IO_IO_IN_DELAY_MAX + 1)) {
1975 * Restore delay chain settings before letting the loop
1976 * in rw_mgr_mem_calibrate_vfifo to retry different
1977 * dqs/ck relationships.
1979 scc_mgr_set_dqs_bus_in_delay(read_group, start_dqs);
1980 if (IO_SHIFT_DQS_EN_WHEN_SHIFT_DQS) {
1981 scc_mgr_set_dqs_en_delay(read_group,
1984 scc_mgr_load_dqs(read_group);
1985 writel(0, &sdr_scc_mgr->update);
1987 debug_cond(DLEVEL == 1, "%s:%d vfifo_center: failed to \
1988 find edge [%u]: %d %d", __func__, __LINE__,
1989 i, left_edge[i], right_edge[i]);
1990 if (use_read_test) {
1991 set_failing_group_stage(read_group *
1992 RW_MGR_MEM_DQ_PER_READ_DQS + i,
1994 CAL_SUBSTAGE_VFIFO_CENTER);
1996 set_failing_group_stage(read_group *
1997 RW_MGR_MEM_DQ_PER_READ_DQS + i,
1998 CAL_STAGE_VFIFO_AFTER_WRITES,
1999 CAL_SUBSTAGE_VFIFO_CENTER);
2005 /* Find middle of window for each DQ bit */
2006 mid_min = left_edge[0] - right_edge[0];
2008 for (i = 1; i < RW_MGR_MEM_DQ_PER_READ_DQS; i++) {
2009 mid = left_edge[i] - right_edge[i];
2010 if (mid < mid_min) {
2017 * -mid_min/2 represents the amount that we need to move DQS.
2018 * If mid_min is odd and positive we'll need to add one to
2019 * make sure the rounding in further calculations is correct
2020 * (always bias to the right), so just add 1 for all positive values.
2025 mid_min = mid_min / 2;
2027 debug_cond(DLEVEL == 1, "%s:%d vfifo_center: mid_min=%d (index=%u)\n",
2028 __func__, __LINE__, mid_min, min_index);
2030 /* Determine the amount we can change DQS (which is -mid_min) */
2031 orig_mid_min = mid_min;
2032 new_dqs = start_dqs - mid_min;
2033 if (new_dqs > IO_DQS_IN_DELAY_MAX)
2034 new_dqs = IO_DQS_IN_DELAY_MAX;
2035 else if (new_dqs < 0)
2038 mid_min = start_dqs - new_dqs;
2039 debug_cond(DLEVEL == 1, "vfifo_center: new mid_min=%d new_dqs=%d\n",
2042 if (IO_SHIFT_DQS_EN_WHEN_SHIFT_DQS) {
2043 if (start_dqs_en - mid_min > IO_DQS_EN_DELAY_MAX)
2044 mid_min += start_dqs_en - mid_min - IO_DQS_EN_DELAY_MAX;
2045 else if (start_dqs_en - mid_min < 0)
2046 mid_min += start_dqs_en - mid_min;
2048 new_dqs = start_dqs - mid_min;
2050 debug_cond(DLEVEL == 1, "vfifo_center: start_dqs=%d start_dqs_en=%d \
2051 new_dqs=%d mid_min=%d\n", start_dqs,
2052 IO_SHIFT_DQS_EN_WHEN_SHIFT_DQS ? start_dqs_en : -1,
2055 /* Initialize data for export structures */
2056 dqs_margin = IO_IO_IN_DELAY_MAX + 1;
2057 dq_margin = IO_IO_IN_DELAY_MAX + 1;
2059 /* add delay to bring centre of all DQ windows to the same "level" */
2060 for (i = 0, p = test_bgn; i < RW_MGR_MEM_DQ_PER_READ_DQS; i++, p++) {
2061 /* Use values before divide by 2 to reduce round off error */
2062 shift_dq = (left_edge[i] - right_edge[i] -
2063 (left_edge[min_index] - right_edge[min_index]))/2 +
2064 (orig_mid_min - mid_min);
2066 debug_cond(DLEVEL == 2, "vfifo_center: before: \
2067 shift_dq[%u]=%d\n", i, shift_dq);
2069 addr = SDR_PHYGRP_SCCGRP_ADDRESS | SCC_MGR_IO_IN_DELAY_OFFSET;
2070 temp_dq_in_delay1 = readl(addr + (p << 2));
2071 temp_dq_in_delay2 = readl(addr + (i << 2));
2073 if (shift_dq + (int32_t)temp_dq_in_delay1 >
2074 (int32_t)IO_IO_IN_DELAY_MAX) {
2075 shift_dq = (int32_t)IO_IO_IN_DELAY_MAX - temp_dq_in_delay2;
2076 } else if (shift_dq + (int32_t)temp_dq_in_delay1 < 0) {
2077 shift_dq = -(int32_t)temp_dq_in_delay1;
2079 debug_cond(DLEVEL == 2, "vfifo_center: after: \
2080 shift_dq[%u]=%d\n", i, shift_dq);
2081 final_dq[i] = temp_dq_in_delay1 + shift_dq;
2082 scc_mgr_set_dq_in_delay(p, final_dq[i]);
2085 debug_cond(DLEVEL == 2, "vfifo_center: margin[%u]=[%d,%d]\n", i,
2086 left_edge[i] - shift_dq + (-mid_min),
2087 right_edge[i] + shift_dq - (-mid_min));
2088 /* To determine values for export structures */
2089 if (left_edge[i] - shift_dq + (-mid_min) < dq_margin)
2090 dq_margin = left_edge[i] - shift_dq + (-mid_min);
2092 if (right_edge[i] + shift_dq - (-mid_min) < dqs_margin)
2093 dqs_margin = right_edge[i] + shift_dq - (-mid_min);
2096 final_dqs = new_dqs;
2097 if (IO_SHIFT_DQS_EN_WHEN_SHIFT_DQS)
2098 final_dqs_en = start_dqs_en - mid_min;
2101 if (IO_SHIFT_DQS_EN_WHEN_SHIFT_DQS) {
2102 scc_mgr_set_dqs_en_delay(read_group, final_dqs_en);
2103 scc_mgr_load_dqs(read_group);
2107 scc_mgr_set_dqs_bus_in_delay(read_group, final_dqs);
2108 scc_mgr_load_dqs(read_group);
2109 debug_cond(DLEVEL == 2, "%s:%d vfifo_center: dq_margin=%d \
2110 dqs_margin=%d", __func__, __LINE__,
2111 dq_margin, dqs_margin);
2114 * Do not remove this line as it makes sure all of our decisions
2115 * have been applied. Apply the update bit.
2117 writel(0, &sdr_scc_mgr->update);
2119 return (dq_margin >= 0) && (dqs_margin >= 0);
2123 * rw_mgr_mem_calibrate_guaranteed_write() - Perform guaranteed write into the device
2124 * @rw_group: Read/Write Group
2125 * @phase: DQ/DQS phase
2127 * Because initially no communication ca be reliably performed with the memory
2128 * device, the sequencer uses a guaranteed write mechanism to write data into
2129 * the memory device.
2131 static int rw_mgr_mem_calibrate_guaranteed_write(const u32 rw_group,
2136 /* Set a particular DQ/DQS phase. */
2137 scc_mgr_set_dqdqs_output_phase_all_ranks(rw_group, phase);
2139 debug_cond(DLEVEL == 1, "%s:%d guaranteed write: g=%u p=%u\n",
2140 __func__, __LINE__, rw_group, phase);
2143 * Altera EMI_RM 2015.05.04 :: Figure 1-25
2144 * Load up the patterns used by read calibration using the
2145 * current DQDQS phase.
2147 rw_mgr_mem_calibrate_read_load_patterns(0, 1);
2149 if (gbl->phy_debug_mode_flags & PHY_DEBUG_DISABLE_GUARANTEED_READ)
2153 * Altera EMI_RM 2015.05.04 :: Figure 1-26
2154 * Back-to-Back reads of the patterns used for calibration.
2156 ret = rw_mgr_mem_calibrate_read_test_patterns(0, rw_group, 1);
2158 debug_cond(DLEVEL == 1,
2159 "%s:%d Guaranteed read test failed: g=%u p=%u\n",
2160 __func__, __LINE__, rw_group, phase);
2165 * rw_mgr_mem_calibrate_dqs_enable_calibration() - DQS Enable Calibration
2166 * @rw_group: Read/Write Group
2167 * @test_bgn: Rank at which the test begins
2169 * DQS enable calibration ensures reliable capture of the DQ signal without
2170 * glitches on the DQS line.
2172 static int rw_mgr_mem_calibrate_dqs_enable_calibration(const u32 rw_group,
2176 * Altera EMI_RM 2015.05.04 :: Figure 1-27
2177 * DQS and DQS Eanble Signal Relationships.
2180 /* We start at zero, so have one less dq to devide among */
2181 const u32 delay_step = IO_IO_IN_DELAY_MAX /
2182 (RW_MGR_MEM_DQ_PER_READ_DQS - 1);
2186 debug("%s:%d (%u,%u)\n", __func__, __LINE__, rw_group, test_bgn);
2188 /* Try different dq_in_delays since the DQ path is shorter than DQS. */
2189 for (r = 0; r < RW_MGR_MEM_NUMBER_OF_RANKS;
2190 r += NUM_RANKS_PER_SHADOW_REG) {
2191 for (i = 0, p = test_bgn, d = 0;
2192 i < RW_MGR_MEM_DQ_PER_READ_DQS;
2193 i++, p++, d += delay_step) {
2194 debug_cond(DLEVEL == 1,
2195 "%s:%d: g=%u r=%u i=%u p=%u d=%u\n",
2196 __func__, __LINE__, rw_group, r, i, p, d);
2198 scc_mgr_set_dq_in_delay(p, d);
2202 writel(0, &sdr_scc_mgr->update);
2206 * Try rw_mgr_mem_calibrate_vfifo_find_dqs_en_phase across different
2207 * dq_in_delay values
2209 found = rw_mgr_mem_calibrate_vfifo_find_dqs_en_phase(rw_group);
2211 debug_cond(DLEVEL == 1,
2212 "%s:%d: g=%u found=%u; Reseting delay chain to zero\n",
2213 __func__, __LINE__, rw_group, found);
2215 for (r = 0; r < RW_MGR_MEM_NUMBER_OF_RANKS;
2216 r += NUM_RANKS_PER_SHADOW_REG) {
2217 scc_mgr_apply_group_dq_in_delay(test_bgn, 0);
2218 writel(0, &sdr_scc_mgr->update);
2229 * rw_mgr_mem_calibrate_dq_dqs_centering() - Centering DQ/DQS
2230 * @rw_group: Read/Write Group
2231 * @test_bgn: Rank at which the test begins
2232 * @use_read_test: Perform a read test
2233 * @update_fom: Update FOM
2235 * The centerin DQ/DQS stage attempts to align DQ and DQS signals on reads
2239 rw_mgr_mem_calibrate_dq_dqs_centering(const u32 rw_group, const u32 test_bgn,
2240 const int use_read_test,
2241 const int update_fom)
2244 int ret, grp_calibrated;
2248 * Altera EMI_RM 2015.05.04 :: Figure 1-28
2249 * Read per-bit deskew can be done on a per shadow register basis.
2252 for (rank_bgn = 0, sr = 0;
2253 rank_bgn < RW_MGR_MEM_NUMBER_OF_RANKS;
2254 rank_bgn += NUM_RANKS_PER_SHADOW_REG, sr++) {
2255 /* Check if this set of ranks should be skipped entirely. */
2256 if (param->skip_shadow_regs[sr])
2259 ret = rw_mgr_mem_calibrate_vfifo_center(rank_bgn, rw_group,
2269 if (!grp_calibrated)
2276 * rw_mgr_mem_calibrate_vfifo() - Calibrate the read valid prediction FIFO
2277 * @rw_group: Read/Write Group
2278 * @test_bgn: Rank at which the test begins
2280 * Stage 1: Calibrate the read valid prediction FIFO.
2282 * This function implements UniPHY calibration Stage 1, as explained in
2283 * detail in Altera EMI_RM 2015.05.04 , "UniPHY Calibration Stages".
2285 * - read valid prediction will consist of finding:
2286 * - DQS enable phase and DQS enable delay (DQS Enable Calibration)
2287 * - DQS input phase and DQS input delay (DQ/DQS Centering)
2288 * - we also do a per-bit deskew on the DQ lines.
2290 static int rw_mgr_mem_calibrate_vfifo(const u32 rw_group, const u32 test_bgn)
2293 uint32_t dtaps_per_ptap;
2294 uint32_t failed_substage;
2298 debug("%s:%d: %u %u\n", __func__, __LINE__, rw_group, test_bgn);
2300 /* Update info for sims */
2301 reg_file_set_group(rw_group);
2302 reg_file_set_stage(CAL_STAGE_VFIFO);
2303 reg_file_set_sub_stage(CAL_SUBSTAGE_GUARANTEED_READ);
2305 failed_substage = CAL_SUBSTAGE_GUARANTEED_READ;
2307 /* USER Determine number of delay taps for each phase tap. */
2308 dtaps_per_ptap = DIV_ROUND_UP(IO_DELAY_PER_OPA_TAP,
2309 IO_DELAY_PER_DQS_EN_DCHAIN_TAP) - 1;
2311 for (d = 0; d <= dtaps_per_ptap; d += 2) {
2313 * In RLDRAMX we may be messing the delay of pins in
2314 * the same write rw_group but outside of the current read
2315 * the rw_group, but that's ok because we haven't calibrated
2319 scc_mgr_apply_group_all_out_delay_add_all_ranks(
2323 for (p = 0; p <= IO_DQDQS_OUT_PHASE_MAX; p++) {
2324 /* 1) Guaranteed Write */
2325 ret = rw_mgr_mem_calibrate_guaranteed_write(rw_group, p);
2329 /* 2) DQS Enable Calibration */
2330 ret = rw_mgr_mem_calibrate_dqs_enable_calibration(rw_group,
2333 failed_substage = CAL_SUBSTAGE_DQS_EN_PHASE;
2337 /* 3) Centering DQ/DQS */
2339 * If doing read after write calibration, do not update
2340 * FOM now. Do it then.
2342 ret = rw_mgr_mem_calibrate_dq_dqs_centering(rw_group,
2345 failed_substage = CAL_SUBSTAGE_VFIFO_CENTER;
2354 /* Calibration Stage 1 failed. */
2355 set_failing_group_stage(rw_group, CAL_STAGE_VFIFO, failed_substage);
2358 /* Calibration Stage 1 completed OK. */
2361 * Reset the delay chains back to zero if they have moved > 1
2362 * (check for > 1 because loop will increase d even when pass in
2366 scc_mgr_zero_group(rw_group, 1);
2371 /* VFIFO Calibration -- Read Deskew Calibration after write deskew */
2372 static uint32_t rw_mgr_mem_calibrate_vfifo_end(uint32_t read_group,
2375 uint32_t rank_bgn, sr;
2376 uint32_t grp_calibrated;
2377 uint32_t write_group;
2379 debug("%s:%d %u %u", __func__, __LINE__, read_group, test_bgn);
2381 /* update info for sims */
2383 reg_file_set_stage(CAL_STAGE_VFIFO_AFTER_WRITES);
2384 reg_file_set_sub_stage(CAL_SUBSTAGE_VFIFO_CENTER);
2386 write_group = read_group;
2388 /* update info for sims */
2389 reg_file_set_group(read_group);
2392 /* Read per-bit deskew can be done on a per shadow register basis */
2393 for (rank_bgn = 0, sr = 0; rank_bgn < RW_MGR_MEM_NUMBER_OF_RANKS;
2394 rank_bgn += NUM_RANKS_PER_SHADOW_REG, ++sr) {
2395 /* Determine if this set of ranks should be skipped entirely */
2396 if (!param->skip_shadow_regs[sr]) {
2397 /* This is the last calibration round, update FOM here */
2398 if (!rw_mgr_mem_calibrate_vfifo_center(rank_bgn,
2409 if (grp_calibrated == 0) {
2410 set_failing_group_stage(write_group,
2411 CAL_STAGE_VFIFO_AFTER_WRITES,
2412 CAL_SUBSTAGE_VFIFO_CENTER);
2419 /* Calibrate LFIFO to find smallest read latency */
2420 static uint32_t rw_mgr_mem_calibrate_lfifo(void)
2425 debug("%s:%d\n", __func__, __LINE__);
2427 /* update info for sims */
2428 reg_file_set_stage(CAL_STAGE_LFIFO);
2429 reg_file_set_sub_stage(CAL_SUBSTAGE_READ_LATENCY);
2431 /* Load up the patterns used by read calibration for all ranks */
2432 rw_mgr_mem_calibrate_read_load_patterns(0, 1);
2436 writel(gbl->curr_read_lat, &phy_mgr_cfg->phy_rlat);
2437 debug_cond(DLEVEL == 2, "%s:%d lfifo: read_lat=%u",
2438 __func__, __LINE__, gbl->curr_read_lat);
2440 if (!rw_mgr_mem_calibrate_read_test_all_ranks(0,
2448 /* reduce read latency and see if things are working */
2450 gbl->curr_read_lat--;
2451 } while (gbl->curr_read_lat > 0);
2453 /* reset the fifos to get pointers to known state */
2455 writel(0, &phy_mgr_cmd->fifo_reset);
2458 /* add a fudge factor to the read latency that was determined */
2459 gbl->curr_read_lat += 2;
2460 writel(gbl->curr_read_lat, &phy_mgr_cfg->phy_rlat);
2461 debug_cond(DLEVEL == 2, "%s:%d lfifo: success: using \
2462 read_lat=%u\n", __func__, __LINE__,
2463 gbl->curr_read_lat);
2466 set_failing_group_stage(0xff, CAL_STAGE_LFIFO,
2467 CAL_SUBSTAGE_READ_LATENCY);
2469 debug_cond(DLEVEL == 2, "%s:%d lfifo: failed at initial \
2470 read_lat=%u\n", __func__, __LINE__,
2471 gbl->curr_read_lat);
2477 * issue write test command.
2478 * two variants are provided. one that just tests a write pattern and
2479 * another that tests datamask functionality.
2481 static void rw_mgr_mem_calibrate_write_test_issue(uint32_t group,
2484 uint32_t mcc_instruction;
2485 uint32_t quick_write_mode = (((STATIC_CALIB_STEPS) & CALIB_SKIP_WRITES) &&
2486 ENABLE_SUPER_QUICK_CALIBRATION);
2487 uint32_t rw_wl_nop_cycles;
2491 * Set counter and jump addresses for the right
2492 * number of NOP cycles.
2493 * The number of supported NOP cycles can range from -1 to infinity
2494 * Three different cases are handled:
2496 * 1. For a number of NOP cycles greater than 0, the RW Mgr looping
2497 * mechanism will be used to insert the right number of NOPs
2499 * 2. For a number of NOP cycles equals to 0, the micro-instruction
2500 * issuing the write command will jump straight to the
2501 * micro-instruction that turns on DQS (for DDRx), or outputs write
2502 * data (for RLD), skipping
2503 * the NOP micro-instruction all together
2505 * 3. A number of NOP cycles equal to -1 indicates that DQS must be
2506 * turned on in the same micro-instruction that issues the write
2507 * command. Then we need
2508 * to directly jump to the micro-instruction that sends out the data
2510 * NOTE: Implementing this mechanism uses 2 RW Mgr jump-counters
2511 * (2 and 3). One jump-counter (0) is used to perform multiple
2512 * write-read operations.
2513 * one counter left to issue this command in "multiple-group" mode
2516 rw_wl_nop_cycles = gbl->rw_wl_nop_cycles;
2518 if (rw_wl_nop_cycles == -1) {
2520 * CNTR 2 - We want to execute the special write operation that
2521 * turns on DQS right away and then skip directly to the
2522 * instruction that sends out the data. We set the counter to a
2523 * large number so that the jump is always taken.
2525 writel(0xFF, &sdr_rw_load_mgr_regs->load_cntr2);
2527 /* CNTR 3 - Not used */
2529 mcc_instruction = RW_MGR_LFSR_WR_RD_DM_BANK_0_WL_1;
2530 writel(RW_MGR_LFSR_WR_RD_DM_BANK_0_DATA,
2531 &sdr_rw_load_jump_mgr_regs->load_jump_add2);
2532 writel(RW_MGR_LFSR_WR_RD_DM_BANK_0_NOP,
2533 &sdr_rw_load_jump_mgr_regs->load_jump_add3);
2535 mcc_instruction = RW_MGR_LFSR_WR_RD_BANK_0_WL_1;
2536 writel(RW_MGR_LFSR_WR_RD_BANK_0_DATA,
2537 &sdr_rw_load_jump_mgr_regs->load_jump_add2);
2538 writel(RW_MGR_LFSR_WR_RD_BANK_0_NOP,
2539 &sdr_rw_load_jump_mgr_regs->load_jump_add3);
2541 } else if (rw_wl_nop_cycles == 0) {
2543 * CNTR 2 - We want to skip the NOP operation and go straight
2544 * to the DQS enable instruction. We set the counter to a large
2545 * number so that the jump is always taken.
2547 writel(0xFF, &sdr_rw_load_mgr_regs->load_cntr2);
2549 /* CNTR 3 - Not used */
2551 mcc_instruction = RW_MGR_LFSR_WR_RD_DM_BANK_0;
2552 writel(RW_MGR_LFSR_WR_RD_DM_BANK_0_DQS,
2553 &sdr_rw_load_jump_mgr_regs->load_jump_add2);
2555 mcc_instruction = RW_MGR_LFSR_WR_RD_BANK_0;
2556 writel(RW_MGR_LFSR_WR_RD_BANK_0_DQS,
2557 &sdr_rw_load_jump_mgr_regs->load_jump_add2);
2561 * CNTR 2 - In this case we want to execute the next instruction
2562 * and NOT take the jump. So we set the counter to 0. The jump
2563 * address doesn't count.
2565 writel(0x0, &sdr_rw_load_mgr_regs->load_cntr2);
2566 writel(0x0, &sdr_rw_load_jump_mgr_regs->load_jump_add2);
2569 * CNTR 3 - Set the nop counter to the number of cycles we
2570 * need to loop for, minus 1.
2572 writel(rw_wl_nop_cycles - 1, &sdr_rw_load_mgr_regs->load_cntr3);
2574 mcc_instruction = RW_MGR_LFSR_WR_RD_DM_BANK_0;
2575 writel(RW_MGR_LFSR_WR_RD_DM_BANK_0_NOP,
2576 &sdr_rw_load_jump_mgr_regs->load_jump_add3);
2578 mcc_instruction = RW_MGR_LFSR_WR_RD_BANK_0;
2579 writel(RW_MGR_LFSR_WR_RD_BANK_0_NOP,
2580 &sdr_rw_load_jump_mgr_regs->load_jump_add3);
2584 writel(0, SDR_PHYGRP_RWMGRGRP_ADDRESS |
2585 RW_MGR_RESET_READ_DATAPATH_OFFSET);
2587 if (quick_write_mode)
2588 writel(0x08, &sdr_rw_load_mgr_regs->load_cntr0);
2590 writel(0x40, &sdr_rw_load_mgr_regs->load_cntr0);
2592 writel(mcc_instruction, &sdr_rw_load_jump_mgr_regs->load_jump_add0);
2595 * CNTR 1 - This is used to ensure enough time elapses
2596 * for read data to come back.
2598 writel(0x30, &sdr_rw_load_mgr_regs->load_cntr1);
2601 writel(RW_MGR_LFSR_WR_RD_DM_BANK_0_WAIT,
2602 &sdr_rw_load_jump_mgr_regs->load_jump_add1);
2604 writel(RW_MGR_LFSR_WR_RD_BANK_0_WAIT,
2605 &sdr_rw_load_jump_mgr_regs->load_jump_add1);
2608 addr = SDR_PHYGRP_RWMGRGRP_ADDRESS | RW_MGR_RUN_SINGLE_GROUP_OFFSET;
2609 writel(mcc_instruction, addr + (group << 2));
2612 /* Test writes, can check for a single bit pass or multiple bit pass */
2613 static uint32_t rw_mgr_mem_calibrate_write_test(uint32_t rank_bgn,
2614 uint32_t write_group, uint32_t use_dm, uint32_t all_correct,
2615 uint32_t *bit_chk, uint32_t all_ranks)
2618 uint32_t correct_mask_vg;
2619 uint32_t tmp_bit_chk;
2621 uint32_t rank_end = all_ranks ? RW_MGR_MEM_NUMBER_OF_RANKS :
2622 (rank_bgn + NUM_RANKS_PER_SHADOW_REG);
2623 uint32_t addr_rw_mgr;
2624 uint32_t base_rw_mgr;
2626 *bit_chk = param->write_correct_mask;
2627 correct_mask_vg = param->write_correct_mask_vg;
2629 for (r = rank_bgn; r < rank_end; r++) {
2630 if (param->skip_ranks[r]) {
2631 /* request to skip the rank */
2636 set_rank_and_odt_mask(r, RW_MGR_ODT_MODE_READ_WRITE);
2639 addr_rw_mgr = SDR_PHYGRP_RWMGRGRP_ADDRESS;
2640 for (vg = RW_MGR_MEM_VIRTUAL_GROUPS_PER_WRITE_DQS-1; ; vg--) {
2641 /* reset the fifos to get pointers to known state */
2642 writel(0, &phy_mgr_cmd->fifo_reset);
2644 tmp_bit_chk = tmp_bit_chk <<
2645 (RW_MGR_MEM_DQ_PER_WRITE_DQS /
2646 RW_MGR_MEM_VIRTUAL_GROUPS_PER_WRITE_DQS);
2647 rw_mgr_mem_calibrate_write_test_issue(write_group *
2648 RW_MGR_MEM_VIRTUAL_GROUPS_PER_WRITE_DQS+vg,
2651 base_rw_mgr = readl(addr_rw_mgr);
2652 tmp_bit_chk = tmp_bit_chk | (correct_mask_vg & ~(base_rw_mgr));
2656 *bit_chk &= tmp_bit_chk;
2660 set_rank_and_odt_mask(0, RW_MGR_ODT_MODE_OFF);
2661 debug_cond(DLEVEL == 2, "write_test(%u,%u,ALL) : %u == \
2662 %u => %lu", write_group, use_dm,
2663 *bit_chk, param->write_correct_mask,
2664 (long unsigned int)(*bit_chk ==
2665 param->write_correct_mask));
2666 return *bit_chk == param->write_correct_mask;
2668 set_rank_and_odt_mask(0, RW_MGR_ODT_MODE_OFF);
2669 debug_cond(DLEVEL == 2, "write_test(%u,%u,ONE) : %u != ",
2670 write_group, use_dm, *bit_chk);
2671 debug_cond(DLEVEL == 2, "%lu" " => %lu", (long unsigned int)0,
2672 (long unsigned int)(*bit_chk != 0));
2673 return *bit_chk != 0x00;
2678 * center all windows. do per-bit-deskew to possibly increase size of
2681 static uint32_t rw_mgr_mem_calibrate_writes_center(uint32_t rank_bgn,
2682 uint32_t write_group, uint32_t test_bgn)
2684 uint32_t i, p, min_index;
2687 * Store these as signed since there are comparisons with
2691 uint32_t sticky_bit_chk;
2692 int32_t left_edge[RW_MGR_MEM_DQ_PER_WRITE_DQS];
2693 int32_t right_edge[RW_MGR_MEM_DQ_PER_WRITE_DQS];
2695 int32_t mid_min, orig_mid_min;
2696 int32_t new_dqs, start_dqs, shift_dq;
2697 int32_t dq_margin, dqs_margin, dm_margin;
2699 uint32_t temp_dq_out1_delay;
2702 debug("%s:%d %u %u", __func__, __LINE__, write_group, test_bgn);
2706 addr = SDR_PHYGRP_SCCGRP_ADDRESS | SCC_MGR_IO_OUT1_DELAY_OFFSET;
2707 start_dqs = readl(addr +
2708 (RW_MGR_MEM_DQ_PER_WRITE_DQS << 2));
2710 /* per-bit deskew */
2713 * set the left and right edge of each bit to an illegal value
2714 * use (IO_IO_OUT1_DELAY_MAX + 1) as an illegal value.
2717 for (i = 0; i < RW_MGR_MEM_DQ_PER_WRITE_DQS; i++) {
2718 left_edge[i] = IO_IO_OUT1_DELAY_MAX + 1;
2719 right_edge[i] = IO_IO_OUT1_DELAY_MAX + 1;
2722 /* Search for the left edge of the window for each bit */
2723 for (d = 0; d <= IO_IO_OUT1_DELAY_MAX; d++) {
2724 scc_mgr_apply_group_dq_out1_delay(write_group, d);
2726 writel(0, &sdr_scc_mgr->update);
2729 * Stop searching when the read test doesn't pass AND when
2730 * we've seen a passing read on every bit.
2732 stop = !rw_mgr_mem_calibrate_write_test(rank_bgn, write_group,
2733 0, PASS_ONE_BIT, &bit_chk, 0);
2734 sticky_bit_chk = sticky_bit_chk | bit_chk;
2735 stop = stop && (sticky_bit_chk == param->write_correct_mask);
2736 debug_cond(DLEVEL == 2, "write_center(left): dtap=%d => %u \
2737 == %u && %u [bit_chk= %u ]\n",
2738 d, sticky_bit_chk, param->write_correct_mask,
2744 for (i = 0; i < RW_MGR_MEM_DQ_PER_WRITE_DQS; i++) {
2747 * Remember a passing test as the
2753 * If a left edge has not been seen
2754 * yet, then a future passing test will
2755 * mark this edge as the right edge.
2758 IO_IO_OUT1_DELAY_MAX + 1) {
2759 right_edge[i] = -(d + 1);
2762 debug_cond(DLEVEL == 2, "write_center[l,d=%d):", d);
2763 debug_cond(DLEVEL == 2, "bit_chk_test=%d left_edge[%u]: %d",
2764 (int)(bit_chk & 1), i, left_edge[i]);
2765 debug_cond(DLEVEL == 2, "right_edge[%u]: %d\n", i,
2767 bit_chk = bit_chk >> 1;
2772 /* Reset DQ delay chains to 0 */
2773 scc_mgr_apply_group_dq_out1_delay(0);
2775 for (i = RW_MGR_MEM_DQ_PER_WRITE_DQS - 1;; i--) {
2776 debug_cond(DLEVEL == 2, "%s:%d write_center: left_edge[%u]: \
2777 %d right_edge[%u]: %d\n", __func__, __LINE__,
2778 i, left_edge[i], i, right_edge[i]);
2781 * Check for cases where we haven't found the left edge,
2782 * which makes our assignment of the the right edge invalid.
2783 * Reset it to the illegal value.
2785 if ((left_edge[i] == IO_IO_OUT1_DELAY_MAX + 1) &&
2786 (right_edge[i] != IO_IO_OUT1_DELAY_MAX + 1)) {
2787 right_edge[i] = IO_IO_OUT1_DELAY_MAX + 1;
2788 debug_cond(DLEVEL == 2, "%s:%d write_center: reset \
2789 right_edge[%u]: %d\n", __func__, __LINE__,
2794 * Reset sticky bit (except for bits where we have
2795 * seen the left edge).
2797 sticky_bit_chk = sticky_bit_chk << 1;
2798 if ((left_edge[i] != IO_IO_OUT1_DELAY_MAX + 1))
2799 sticky_bit_chk = sticky_bit_chk | 1;
2805 /* Search for the right edge of the window for each bit */
2806 for (d = 0; d <= IO_IO_OUT1_DELAY_MAX - start_dqs; d++) {
2807 scc_mgr_apply_group_dqs_io_and_oct_out1(write_group,
2810 writel(0, &sdr_scc_mgr->update);
2813 * Stop searching when the read test doesn't pass AND when
2814 * we've seen a passing read on every bit.
2816 stop = !rw_mgr_mem_calibrate_write_test(rank_bgn, write_group,
2817 0, PASS_ONE_BIT, &bit_chk, 0);
2819 sticky_bit_chk = sticky_bit_chk | bit_chk;
2820 stop = stop && (sticky_bit_chk == param->write_correct_mask);
2822 debug_cond(DLEVEL == 2, "write_center (right): dtap=%u => %u == \
2823 %u && %u\n", d, sticky_bit_chk,
2824 param->write_correct_mask, stop);
2828 for (i = 0; i < RW_MGR_MEM_DQ_PER_WRITE_DQS;
2830 /* d = 0 failed, but it passed when
2831 testing the left edge, so it must be
2832 marginal, set it to -1 */
2833 if (right_edge[i] ==
2834 IO_IO_OUT1_DELAY_MAX + 1 &&
2836 IO_IO_OUT1_DELAY_MAX + 1) {
2843 for (i = 0; i < RW_MGR_MEM_DQ_PER_WRITE_DQS; i++) {
2846 * Remember a passing test as
2853 * If a right edge has not
2854 * been seen yet, then a future
2855 * passing test will mark this
2856 * edge as the left edge.
2858 if (right_edge[i] ==
2859 IO_IO_OUT1_DELAY_MAX + 1)
2860 left_edge[i] = -(d + 1);
2863 * d = 0 failed, but it passed
2864 * when testing the left edge,
2865 * so it must be marginal, set
2868 if (right_edge[i] ==
2869 IO_IO_OUT1_DELAY_MAX + 1 &&
2871 IO_IO_OUT1_DELAY_MAX + 1)
2874 * If a right edge has not been
2875 * seen yet, then a future
2876 * passing test will mark this
2877 * edge as the left edge.
2879 else if (right_edge[i] ==
2880 IO_IO_OUT1_DELAY_MAX +
2882 left_edge[i] = -(d + 1);
2885 debug_cond(DLEVEL == 2, "write_center[r,d=%d):", d);
2886 debug_cond(DLEVEL == 2, "bit_chk_test=%d left_edge[%u]: %d",
2887 (int)(bit_chk & 1), i, left_edge[i]);
2888 debug_cond(DLEVEL == 2, "right_edge[%u]: %d\n", i,
2890 bit_chk = bit_chk >> 1;
2895 /* Check that all bits have a window */
2896 for (i = 0; i < RW_MGR_MEM_DQ_PER_WRITE_DQS; i++) {
2897 debug_cond(DLEVEL == 2, "%s:%d write_center: left_edge[%u]: \
2898 %d right_edge[%u]: %d", __func__, __LINE__,
2899 i, left_edge[i], i, right_edge[i]);
2900 if ((left_edge[i] == IO_IO_OUT1_DELAY_MAX + 1) ||
2901 (right_edge[i] == IO_IO_OUT1_DELAY_MAX + 1)) {
2902 set_failing_group_stage(test_bgn + i,
2904 CAL_SUBSTAGE_WRITES_CENTER);
2909 /* Find middle of window for each DQ bit */
2910 mid_min = left_edge[0] - right_edge[0];
2912 for (i = 1; i < RW_MGR_MEM_DQ_PER_WRITE_DQS; i++) {
2913 mid = left_edge[i] - right_edge[i];
2914 if (mid < mid_min) {
2921 * -mid_min/2 represents the amount that we need to move DQS.
2922 * If mid_min is odd and positive we'll need to add one to
2923 * make sure the rounding in further calculations is correct
2924 * (always bias to the right), so just add 1 for all positive values.
2928 mid_min = mid_min / 2;
2929 debug_cond(DLEVEL == 1, "%s:%d write_center: mid_min=%d\n", __func__,
2932 /* Determine the amount we can change DQS (which is -mid_min) */
2933 orig_mid_min = mid_min;
2934 new_dqs = start_dqs;
2936 debug_cond(DLEVEL == 1, "%s:%d write_center: start_dqs=%d new_dqs=%d \
2937 mid_min=%d\n", __func__, __LINE__, start_dqs, new_dqs, mid_min);
2938 /* Initialize data for export structures */
2939 dqs_margin = IO_IO_OUT1_DELAY_MAX + 1;
2940 dq_margin = IO_IO_OUT1_DELAY_MAX + 1;
2942 /* add delay to bring centre of all DQ windows to the same "level" */
2943 for (i = 0, p = test_bgn; i < RW_MGR_MEM_DQ_PER_WRITE_DQS; i++, p++) {
2944 /* Use values before divide by 2 to reduce round off error */
2945 shift_dq = (left_edge[i] - right_edge[i] -
2946 (left_edge[min_index] - right_edge[min_index]))/2 +
2947 (orig_mid_min - mid_min);
2949 debug_cond(DLEVEL == 2, "%s:%d write_center: before: shift_dq \
2950 [%u]=%d\n", __func__, __LINE__, i, shift_dq);
2952 addr = SDR_PHYGRP_SCCGRP_ADDRESS | SCC_MGR_IO_OUT1_DELAY_OFFSET;
2953 temp_dq_out1_delay = readl(addr + (i << 2));
2954 if (shift_dq + (int32_t)temp_dq_out1_delay >
2955 (int32_t)IO_IO_OUT1_DELAY_MAX) {
2956 shift_dq = (int32_t)IO_IO_OUT1_DELAY_MAX - temp_dq_out1_delay;
2957 } else if (shift_dq + (int32_t)temp_dq_out1_delay < 0) {
2958 shift_dq = -(int32_t)temp_dq_out1_delay;
2960 debug_cond(DLEVEL == 2, "write_center: after: shift_dq[%u]=%d\n",
2962 scc_mgr_set_dq_out1_delay(i, temp_dq_out1_delay + shift_dq);
2965 debug_cond(DLEVEL == 2, "write_center: margin[%u]=[%d,%d]\n", i,
2966 left_edge[i] - shift_dq + (-mid_min),
2967 right_edge[i] + shift_dq - (-mid_min));
2968 /* To determine values for export structures */
2969 if (left_edge[i] - shift_dq + (-mid_min) < dq_margin)
2970 dq_margin = left_edge[i] - shift_dq + (-mid_min);
2972 if (right_edge[i] + shift_dq - (-mid_min) < dqs_margin)
2973 dqs_margin = right_edge[i] + shift_dq - (-mid_min);
2977 scc_mgr_apply_group_dqs_io_and_oct_out1(write_group, new_dqs);
2978 writel(0, &sdr_scc_mgr->update);
2981 debug_cond(DLEVEL == 2, "%s:%d write_center: DM\n", __func__, __LINE__);
2984 * set the left and right edge of each bit to an illegal value,
2985 * use (IO_IO_OUT1_DELAY_MAX + 1) as an illegal value,
2987 left_edge[0] = IO_IO_OUT1_DELAY_MAX + 1;
2988 right_edge[0] = IO_IO_OUT1_DELAY_MAX + 1;
2989 int32_t bgn_curr = IO_IO_OUT1_DELAY_MAX + 1;
2990 int32_t end_curr = IO_IO_OUT1_DELAY_MAX + 1;
2991 int32_t bgn_best = IO_IO_OUT1_DELAY_MAX + 1;
2992 int32_t end_best = IO_IO_OUT1_DELAY_MAX + 1;
2993 int32_t win_best = 0;
2995 /* Search for the/part of the window with DM shift */
2996 for (d = IO_IO_OUT1_DELAY_MAX; d >= 0; d -= DELTA_D) {
2997 scc_mgr_apply_group_dm_out1_delay(d);
2998 writel(0, &sdr_scc_mgr->update);
3000 if (rw_mgr_mem_calibrate_write_test(rank_bgn, write_group, 1,
3001 PASS_ALL_BITS, &bit_chk,
3003 /* USE Set current end of the window */
3006 * If a starting edge of our window has not been seen
3007 * this is our current start of the DM window.
3009 if (bgn_curr == IO_IO_OUT1_DELAY_MAX + 1)
3013 * If current window is bigger than best seen.
3014 * Set best seen to be current window.
3016 if ((end_curr-bgn_curr+1) > win_best) {
3017 win_best = end_curr-bgn_curr+1;
3018 bgn_best = bgn_curr;
3019 end_best = end_curr;
3022 /* We just saw a failing test. Reset temp edge */
3023 bgn_curr = IO_IO_OUT1_DELAY_MAX + 1;
3024 end_curr = IO_IO_OUT1_DELAY_MAX + 1;
3029 /* Reset DM delay chains to 0 */
3030 scc_mgr_apply_group_dm_out1_delay(0);
3033 * Check to see if the current window nudges up aganist 0 delay.
3034 * If so we need to continue the search by shifting DQS otherwise DQS
3035 * search begins as a new search. */
3036 if (end_curr != 0) {
3037 bgn_curr = IO_IO_OUT1_DELAY_MAX + 1;
3038 end_curr = IO_IO_OUT1_DELAY_MAX + 1;
3041 /* Search for the/part of the window with DQS shifts */
3042 for (d = 0; d <= IO_IO_OUT1_DELAY_MAX - new_dqs; d += DELTA_D) {
3044 * Note: This only shifts DQS, so are we limiting ourselve to
3045 * width of DQ unnecessarily.
3047 scc_mgr_apply_group_dqs_io_and_oct_out1(write_group,
3050 writel(0, &sdr_scc_mgr->update);
3051 if (rw_mgr_mem_calibrate_write_test(rank_bgn, write_group, 1,
3052 PASS_ALL_BITS, &bit_chk,
3054 /* USE Set current end of the window */
3057 * If a beginning edge of our window has not been seen
3058 * this is our current begin of the DM window.
3060 if (bgn_curr == IO_IO_OUT1_DELAY_MAX + 1)
3064 * If current window is bigger than best seen. Set best
3065 * seen to be current window.
3067 if ((end_curr-bgn_curr+1) > win_best) {
3068 win_best = end_curr-bgn_curr+1;
3069 bgn_best = bgn_curr;
3070 end_best = end_curr;
3073 /* We just saw a failing test. Reset temp edge */
3074 bgn_curr = IO_IO_OUT1_DELAY_MAX + 1;
3075 end_curr = IO_IO_OUT1_DELAY_MAX + 1;
3077 /* Early exit optimization: if ther remaining delay
3078 chain space is less than already seen largest window
3081 (IO_IO_OUT1_DELAY_MAX - new_dqs - d)) {
3087 /* assign left and right edge for cal and reporting; */
3088 left_edge[0] = -1*bgn_best;
3089 right_edge[0] = end_best;
3091 debug_cond(DLEVEL == 2, "%s:%d dm_calib: left=%d right=%d\n", __func__,
3092 __LINE__, left_edge[0], right_edge[0]);
3094 /* Move DQS (back to orig) */
3095 scc_mgr_apply_group_dqs_io_and_oct_out1(write_group, new_dqs);
3099 /* Find middle of window for the DM bit */
3100 mid = (left_edge[0] - right_edge[0]) / 2;
3102 /* only move right, since we are not moving DQS/DQ */
3106 /* dm_marign should fail if we never find a window */
3110 dm_margin = left_edge[0] - mid;
3112 scc_mgr_apply_group_dm_out1_delay(mid);
3113 writel(0, &sdr_scc_mgr->update);
3115 debug_cond(DLEVEL == 2, "%s:%d dm_calib: left=%d right=%d mid=%d \
3116 dm_margin=%d\n", __func__, __LINE__, left_edge[0],
3117 right_edge[0], mid, dm_margin);
3119 gbl->fom_out += dq_margin + dqs_margin;
3121 debug_cond(DLEVEL == 2, "%s:%d write_center: dq_margin=%d \
3122 dqs_margin=%d dm_margin=%d\n", __func__, __LINE__,
3123 dq_margin, dqs_margin, dm_margin);
3126 * Do not remove this line as it makes sure all of our
3127 * decisions have been applied.
3129 writel(0, &sdr_scc_mgr->update);
3130 return (dq_margin >= 0) && (dqs_margin >= 0) && (dm_margin >= 0);
3133 /* calibrate the write operations */
3134 static uint32_t rw_mgr_mem_calibrate_writes(uint32_t rank_bgn, uint32_t g,
3137 /* update info for sims */
3138 debug("%s:%d %u %u\n", __func__, __LINE__, g, test_bgn);
3140 reg_file_set_stage(CAL_STAGE_WRITES);
3141 reg_file_set_sub_stage(CAL_SUBSTAGE_WRITES_CENTER);
3143 reg_file_set_group(g);
3145 if (!rw_mgr_mem_calibrate_writes_center(rank_bgn, g, test_bgn)) {
3146 set_failing_group_stage(g, CAL_STAGE_WRITES,
3147 CAL_SUBSTAGE_WRITES_CENTER);
3155 * mem_precharge_and_activate() - Precharge all banks and activate
3157 * Precharge all banks and activate row 0 in bank "000..." and bank "111...".
3159 static void mem_precharge_and_activate(void)
3163 for (r = 0; r < RW_MGR_MEM_NUMBER_OF_RANKS; r++) {
3164 /* Test if the rank should be skipped. */
3165 if (param->skip_ranks[r])
3169 set_rank_and_odt_mask(r, RW_MGR_ODT_MODE_OFF);
3171 /* Precharge all banks. */
3172 writel(RW_MGR_PRECHARGE_ALL, SDR_PHYGRP_RWMGRGRP_ADDRESS |
3173 RW_MGR_RUN_SINGLE_GROUP_OFFSET);
3175 writel(0x0F, &sdr_rw_load_mgr_regs->load_cntr0);
3176 writel(RW_MGR_ACTIVATE_0_AND_1_WAIT1,
3177 &sdr_rw_load_jump_mgr_regs->load_jump_add0);
3179 writel(0x0F, &sdr_rw_load_mgr_regs->load_cntr1);
3180 writel(RW_MGR_ACTIVATE_0_AND_1_WAIT2,
3181 &sdr_rw_load_jump_mgr_regs->load_jump_add1);
3183 /* Activate rows. */
3184 writel(RW_MGR_ACTIVATE_0_AND_1, SDR_PHYGRP_RWMGRGRP_ADDRESS |
3185 RW_MGR_RUN_SINGLE_GROUP_OFFSET);
3190 * mem_init_latency() - Configure memory RLAT and WLAT settings
3192 * Configure memory RLAT and WLAT parameters.
3194 static void mem_init_latency(void)
3197 * For AV/CV, LFIFO is hardened and always runs at full rate
3198 * so max latency in AFI clocks, used here, is correspondingly
3201 const u32 max_latency = (1 << MAX_LATENCY_COUNT_WIDTH) - 1;
3204 debug("%s:%d\n", __func__, __LINE__);
3207 * Read in write latency.
3208 * WL for Hard PHY does not include additive latency.
3210 wlat = readl(&data_mgr->t_wl_add);
3211 wlat += readl(&data_mgr->mem_t_add);
3213 gbl->rw_wl_nop_cycles = wlat - 1;
3215 /* Read in readl latency. */
3216 rlat = readl(&data_mgr->t_rl_add);
3218 /* Set a pretty high read latency initially. */
3219 gbl->curr_read_lat = rlat + 16;
3220 if (gbl->curr_read_lat > max_latency)
3221 gbl->curr_read_lat = max_latency;
3223 writel(gbl->curr_read_lat, &phy_mgr_cfg->phy_rlat);
3225 /* Advertise write latency. */
3226 writel(wlat, &phy_mgr_cfg->afi_wlat);
3230 * @mem_skip_calibrate() - Set VFIFO and LFIFO to instant-on settings
3232 * Set VFIFO and LFIFO to instant-on settings in skip calibration mode.
3234 static void mem_skip_calibrate(void)
3236 uint32_t vfifo_offset;
3239 debug("%s:%d\n", __func__, __LINE__);
3240 /* Need to update every shadow register set used by the interface */
3241 for (r = 0; r < RW_MGR_MEM_NUMBER_OF_RANKS;
3242 r += NUM_RANKS_PER_SHADOW_REG) {
3244 * Set output phase alignment settings appropriate for
3247 for (i = 0; i < RW_MGR_MEM_IF_READ_DQS_WIDTH; i++) {
3248 scc_mgr_set_dqs_en_phase(i, 0);
3249 #if IO_DLL_CHAIN_LENGTH == 6
3250 scc_mgr_set_dqdqs_output_phase(i, 6);
3252 scc_mgr_set_dqdqs_output_phase(i, 7);
3257 * Write data arrives to the I/O two cycles before write
3258 * latency is reached (720 deg).
3259 * -> due to bit-slip in a/c bus
3260 * -> to allow board skew where dqs is longer than ck
3261 * -> how often can this happen!?
3262 * -> can claim back some ptaps for high freq
3263 * support if we can relax this, but i digress...
3265 * The write_clk leads mem_ck by 90 deg
3266 * The minimum ptap of the OPA is 180 deg
3267 * Each ptap has (360 / IO_DLL_CHAIN_LENGH) deg of delay
3268 * The write_clk is always delayed by 2 ptaps
3270 * Hence, to make DQS aligned to CK, we need to delay
3272 * (720 - 90 - 180 - 2 * (360 / IO_DLL_CHAIN_LENGTH))
3274 * Dividing the above by (360 / IO_DLL_CHAIN_LENGTH)
3275 * gives us the number of ptaps, which simplies to:
3277 * (1.25 * IO_DLL_CHAIN_LENGTH - 2)
3279 scc_mgr_set_dqdqs_output_phase(i,
3280 1.25 * IO_DLL_CHAIN_LENGTH - 2);
3282 writel(0xff, &sdr_scc_mgr->dqs_ena);
3283 writel(0xff, &sdr_scc_mgr->dqs_io_ena);
3285 for (i = 0; i < RW_MGR_MEM_IF_WRITE_DQS_WIDTH; i++) {
3286 writel(i, SDR_PHYGRP_SCCGRP_ADDRESS |
3287 SCC_MGR_GROUP_COUNTER_OFFSET);
3289 writel(0xff, &sdr_scc_mgr->dq_ena);
3290 writel(0xff, &sdr_scc_mgr->dm_ena);
3291 writel(0, &sdr_scc_mgr->update);
3294 /* Compensate for simulation model behaviour */
3295 for (i = 0; i < RW_MGR_MEM_IF_READ_DQS_WIDTH; i++) {
3296 scc_mgr_set_dqs_bus_in_delay(i, 10);
3297 scc_mgr_load_dqs(i);
3299 writel(0, &sdr_scc_mgr->update);
3302 * ArriaV has hard FIFOs that can only be initialized by incrementing
3305 vfifo_offset = CALIB_VFIFO_OFFSET;
3306 for (j = 0; j < vfifo_offset; j++)
3307 writel(0xff, &phy_mgr_cmd->inc_vfifo_hard_phy);
3308 writel(0, &phy_mgr_cmd->fifo_reset);
3311 * For Arria V and Cyclone V with hard LFIFO, we get the skip-cal
3312 * setting from generation-time constant.
3314 gbl->curr_read_lat = CALIB_LFIFO_OFFSET;
3315 writel(gbl->curr_read_lat, &phy_mgr_cfg->phy_rlat);
3319 * mem_calibrate() - Memory calibration entry point.
3321 * Perform memory calibration.
3323 static uint32_t mem_calibrate(void)
3326 uint32_t rank_bgn, sr;
3327 uint32_t write_group, write_test_bgn;
3328 uint32_t read_group, read_test_bgn;
3329 uint32_t run_groups, current_run;
3330 uint32_t failing_groups = 0;
3331 uint32_t group_failed = 0;
3333 const u32 rwdqs_ratio = RW_MGR_MEM_IF_READ_DQS_WIDTH /
3334 RW_MGR_MEM_IF_WRITE_DQS_WIDTH;
3336 debug("%s:%d\n", __func__, __LINE__);
3338 /* Initialize the data settings */
3339 gbl->error_substage = CAL_SUBSTAGE_NIL;
3340 gbl->error_stage = CAL_STAGE_NIL;
3341 gbl->error_group = 0xff;
3345 /* Initialize WLAT and RLAT. */
3348 /* Initialize bit slips. */
3349 mem_precharge_and_activate();
3351 for (i = 0; i < RW_MGR_MEM_IF_READ_DQS_WIDTH; i++) {
3352 writel(i, SDR_PHYGRP_SCCGRP_ADDRESS |
3353 SCC_MGR_GROUP_COUNTER_OFFSET);
3354 /* Only needed once to set all groups, pins, DQ, DQS, DM. */
3356 scc_mgr_set_hhp_extras();
3358 scc_set_bypass_mode(i);
3361 /* Calibration is skipped. */
3362 if ((dyn_calib_steps & CALIB_SKIP_ALL) == CALIB_SKIP_ALL) {
3364 * Set VFIFO and LFIFO to instant-on settings in skip
3367 mem_skip_calibrate();
3370 * Do not remove this line as it makes sure all of our
3371 * decisions have been applied.
3373 writel(0, &sdr_scc_mgr->update);
3377 /* Calibration is not skipped. */
3378 for (i = 0; i < NUM_CALIB_REPEAT; i++) {
3380 * Zero all delay chain/phase settings for all
3381 * groups and all shadow register sets.
3385 run_groups = ~param->skip_groups;
3387 for (write_group = 0, write_test_bgn = 0; write_group
3388 < RW_MGR_MEM_IF_WRITE_DQS_WIDTH; write_group++,
3389 write_test_bgn += RW_MGR_MEM_DQ_PER_WRITE_DQS) {
3391 /* Initialize the group failure */
3394 current_run = run_groups & ((1 <<
3395 RW_MGR_NUM_DQS_PER_WRITE_GROUP) - 1);
3396 run_groups = run_groups >>
3397 RW_MGR_NUM_DQS_PER_WRITE_GROUP;
3399 if (current_run == 0)
3402 writel(write_group, SDR_PHYGRP_SCCGRP_ADDRESS |
3403 SCC_MGR_GROUP_COUNTER_OFFSET);
3404 scc_mgr_zero_group(write_group, 0);
3406 for (read_group = write_group * rwdqs_ratio,
3408 read_group < (write_group + 1) * rwdqs_ratio;
3410 read_test_bgn += RW_MGR_MEM_DQ_PER_READ_DQS) {
3411 if (STATIC_CALIB_STEPS & CALIB_SKIP_VFIFO)
3414 /* Calibrate the VFIFO */
3415 if (rw_mgr_mem_calibrate_vfifo(read_group,
3419 if (!(gbl->phy_debug_mode_flags & PHY_DEBUG_SWEEP_ALL_GROUPS))
3422 /* The group failed, we're done. */
3426 /* Calibrate the output side */
3427 for (rank_bgn = 0, sr = 0;
3428 rank_bgn < RW_MGR_MEM_NUMBER_OF_RANKS;
3429 rank_bgn += NUM_RANKS_PER_SHADOW_REG, sr++) {
3430 if (STATIC_CALIB_STEPS & CALIB_SKIP_WRITES)
3433 /* Not needed in quick mode! */
3434 if (STATIC_CALIB_STEPS & CALIB_SKIP_DELAY_SWEEPS)
3438 * Determine if this set of ranks
3439 * should be skipped entirely.
3441 if (param->skip_shadow_regs[sr])
3444 /* Calibrate WRITEs */
3445 if (rw_mgr_mem_calibrate_writes(rank_bgn,
3446 write_group, write_test_bgn))
3450 if (!(gbl->phy_debug_mode_flags & PHY_DEBUG_SWEEP_ALL_GROUPS))
3454 /* Some group failed, we're done. */
3458 for (read_group = write_group * rwdqs_ratio,
3460 read_group < (write_group + 1) * rwdqs_ratio;
3462 read_test_bgn += RW_MGR_MEM_DQ_PER_READ_DQS) {
3463 if (STATIC_CALIB_STEPS & CALIB_SKIP_WRITES)
3466 if (rw_mgr_mem_calibrate_vfifo_end(read_group,
3470 if (!(gbl->phy_debug_mode_flags & PHY_DEBUG_SWEEP_ALL_GROUPS))
3473 /* The group failed, we're done. */
3477 /* No group failed, continue as usual. */
3480 grp_failed: /* A group failed, increment the counter. */
3485 * USER If there are any failing groups then report
3488 if (failing_groups != 0)
3491 if (STATIC_CALIB_STEPS & CALIB_SKIP_LFIFO)
3495 * If we're skipping groups as part of debug,
3496 * don't calibrate LFIFO.
3498 if (param->skip_groups != 0)
3501 /* Calibrate the LFIFO */
3502 if (!rw_mgr_mem_calibrate_lfifo())
3507 * Do not remove this line as it makes sure all of our decisions
3508 * have been applied.
3510 writel(0, &sdr_scc_mgr->update);
3515 * run_mem_calibrate() - Perform memory calibration
3517 * This function triggers the entire memory calibration procedure.
3519 static int run_mem_calibrate(void)
3523 debug("%s:%d\n", __func__, __LINE__);
3525 /* Reset pass/fail status shown on afi_cal_success/fail */
3526 writel(PHY_MGR_CAL_RESET, &phy_mgr_cfg->cal_status);
3528 /* Stop tracking manager. */
3529 clrbits_le32(&sdr_ctrl->ctrl_cfg, 1 << 22);
3531 phy_mgr_initialize();
3532 rw_mgr_mem_initialize();
3534 /* Perform the actual memory calibration. */
3535 pass = mem_calibrate();
3537 mem_precharge_and_activate();
3538 writel(0, &phy_mgr_cmd->fifo_reset);
3541 rw_mgr_mem_handoff();
3543 * In Hard PHY this is a 2-bit control:
3545 * 1: DDIO Mux Select
3547 writel(0x2, &phy_mgr_cfg->mux_sel);
3549 /* Start tracking manager. */
3550 setbits_le32(&sdr_ctrl->ctrl_cfg, 1 << 22);
3556 * debug_mem_calibrate() - Report result of memory calibration
3557 * @pass: Value indicating whether calibration passed or failed
3559 * This function reports the results of the memory calibration
3560 * and writes debug information into the register file.
3562 static void debug_mem_calibrate(int pass)
3564 uint32_t debug_info;
3567 printf("%s: CALIBRATION PASSED\n", __FILE__);
3572 if (gbl->fom_in > 0xff)
3575 if (gbl->fom_out > 0xff)
3576 gbl->fom_out = 0xff;
3578 /* Update the FOM in the register file */
3579 debug_info = gbl->fom_in;
3580 debug_info |= gbl->fom_out << 8;
3581 writel(debug_info, &sdr_reg_file->fom);
3583 writel(debug_info, &phy_mgr_cfg->cal_debug_info);
3584 writel(PHY_MGR_CAL_SUCCESS, &phy_mgr_cfg->cal_status);
3586 printf("%s: CALIBRATION FAILED\n", __FILE__);
3588 debug_info = gbl->error_stage;
3589 debug_info |= gbl->error_substage << 8;
3590 debug_info |= gbl->error_group << 16;
3592 writel(debug_info, &sdr_reg_file->failing_stage);
3593 writel(debug_info, &phy_mgr_cfg->cal_debug_info);
3594 writel(PHY_MGR_CAL_FAIL, &phy_mgr_cfg->cal_status);
3596 /* Update the failing group/stage in the register file */
3597 debug_info = gbl->error_stage;
3598 debug_info |= gbl->error_substage << 8;
3599 debug_info |= gbl->error_group << 16;
3600 writel(debug_info, &sdr_reg_file->failing_stage);
3603 printf("%s: Calibration complete\n", __FILE__);
3607 * hc_initialize_rom_data() - Initialize ROM data
3609 * Initialize ROM data.
3611 static void hc_initialize_rom_data(void)
3615 addr = SDR_PHYGRP_RWMGRGRP_ADDRESS | RW_MGR_INST_ROM_WRITE_OFFSET;
3616 for (i = 0; i < ARRAY_SIZE(inst_rom_init); i++)
3617 writel(inst_rom_init[i], addr + (i << 2));
3619 addr = SDR_PHYGRP_RWMGRGRP_ADDRESS | RW_MGR_AC_ROM_WRITE_OFFSET;
3620 for (i = 0; i < ARRAY_SIZE(ac_rom_init); i++)
3621 writel(ac_rom_init[i], addr + (i << 2));
3625 * initialize_reg_file() - Initialize SDR register file
3627 * Initialize SDR register file.
3629 static void initialize_reg_file(void)
3631 /* Initialize the register file with the correct data */
3632 writel(REG_FILE_INIT_SEQ_SIGNATURE, &sdr_reg_file->signature);
3633 writel(0, &sdr_reg_file->debug_data_addr);
3634 writel(0, &sdr_reg_file->cur_stage);
3635 writel(0, &sdr_reg_file->fom);
3636 writel(0, &sdr_reg_file->failing_stage);
3637 writel(0, &sdr_reg_file->debug1);
3638 writel(0, &sdr_reg_file->debug2);
3642 * initialize_hps_phy() - Initialize HPS PHY
3644 * Initialize HPS PHY.
3646 static void initialize_hps_phy(void)
3650 * Tracking also gets configured here because it's in the
3653 uint32_t trk_sample_count = 7500;
3654 uint32_t trk_long_idle_sample_count = (10 << 16) | 100;
3656 * Format is number of outer loops in the 16 MSB, sample
3661 reg |= SDR_CTRLGRP_PHYCTRL_PHYCTRL_0_ACDELAYEN_SET(2);
3662 reg |= SDR_CTRLGRP_PHYCTRL_PHYCTRL_0_DQDELAYEN_SET(1);
3663 reg |= SDR_CTRLGRP_PHYCTRL_PHYCTRL_0_DQSDELAYEN_SET(1);
3664 reg |= SDR_CTRLGRP_PHYCTRL_PHYCTRL_0_DQSLOGICDELAYEN_SET(1);
3665 reg |= SDR_CTRLGRP_PHYCTRL_PHYCTRL_0_RESETDELAYEN_SET(0);
3666 reg |= SDR_CTRLGRP_PHYCTRL_PHYCTRL_0_LPDDRDIS_SET(1);
3668 * This field selects the intrinsic latency to RDATA_EN/FULL path.
3669 * 00-bypass, 01- add 5 cycles, 10- add 10 cycles, 11- add 15 cycles.
3671 reg |= SDR_CTRLGRP_PHYCTRL_PHYCTRL_0_ADDLATSEL_SET(0);
3672 reg |= SDR_CTRLGRP_PHYCTRL_PHYCTRL_0_SAMPLECOUNT_19_0_SET(
3674 writel(reg, &sdr_ctrl->phy_ctrl0);
3677 reg |= SDR_CTRLGRP_PHYCTRL_PHYCTRL_1_SAMPLECOUNT_31_20_SET(
3679 SDR_CTRLGRP_PHYCTRL_PHYCTRL_0_SAMPLECOUNT_19_0_WIDTH);
3680 reg |= SDR_CTRLGRP_PHYCTRL_PHYCTRL_1_LONGIDLESAMPLECOUNT_19_0_SET(
3681 trk_long_idle_sample_count);
3682 writel(reg, &sdr_ctrl->phy_ctrl1);
3685 reg |= SDR_CTRLGRP_PHYCTRL_PHYCTRL_2_LONGIDLESAMPLECOUNT_31_20_SET(
3686 trk_long_idle_sample_count >>
3687 SDR_CTRLGRP_PHYCTRL_PHYCTRL_1_LONGIDLESAMPLECOUNT_19_0_WIDTH);
3688 writel(reg, &sdr_ctrl->phy_ctrl2);
3692 * initialize_tracking() - Initialize tracking
3694 * Initialize the register file with usable initial data.
3696 static void initialize_tracking(void)
3699 * Initialize the register file with the correct data.
3700 * Compute usable version of value in case we skip full
3701 * computation later.
3703 writel(DIV_ROUND_UP(IO_DELAY_PER_OPA_TAP, IO_DELAY_PER_DCHAIN_TAP) - 1,
3704 &sdr_reg_file->dtaps_per_ptap);
3706 /* trk_sample_count */
3707 writel(7500, &sdr_reg_file->trk_sample_count);
3709 /* longidle outer loop [15:0] */
3710 writel((10 << 16) | (100 << 0), &sdr_reg_file->trk_longidle);
3713 * longidle sample count [31:24]
3714 * trfc, worst case of 933Mhz 4Gb [23:16]
3715 * trcd, worst case [15:8]
3718 writel((243 << 24) | (14 << 16) | (10 << 8) | (4 << 0),
3719 &sdr_reg_file->delays);
3722 writel((RW_MGR_IDLE << 24) | (RW_MGR_ACTIVATE_1 << 16) |
3723 (RW_MGR_SGLE_READ << 8) | (RW_MGR_PRECHARGE_ALL << 0),
3724 &sdr_reg_file->trk_rw_mgr_addr);
3726 writel(RW_MGR_MEM_IF_READ_DQS_WIDTH,
3727 &sdr_reg_file->trk_read_dqs_width);
3730 writel((RW_MGR_REFRESH_ALL << 24) | (1000 << 0),
3731 &sdr_reg_file->trk_rfsh);
3734 int sdram_calibration_full(void)
3736 struct param_type my_param;
3737 struct gbl_type my_gbl;
3740 memset(&my_param, 0, sizeof(my_param));
3741 memset(&my_gbl, 0, sizeof(my_gbl));
3746 /* Set the calibration enabled by default */
3747 gbl->phy_debug_mode_flags |= PHY_DEBUG_ENABLE_CAL_RPT;
3749 * Only sweep all groups (regardless of fail state) by default
3750 * Set enabled read test by default.
3752 #if DISABLE_GUARANTEED_READ
3753 gbl->phy_debug_mode_flags |= PHY_DEBUG_DISABLE_GUARANTEED_READ;
3755 /* Initialize the register file */
3756 initialize_reg_file();
3758 /* Initialize any PHY CSR */
3759 initialize_hps_phy();
3761 scc_mgr_initialize();
3763 initialize_tracking();
3765 printf("%s: Preparing to start memory calibration\n", __FILE__);
3767 debug("%s:%d\n", __func__, __LINE__);
3768 debug_cond(DLEVEL == 1,
3769 "DDR3 FULL_RATE ranks=%u cs/dimm=%u dq/dqs=%u,%u vg/dqs=%u,%u ",
3770 RW_MGR_MEM_NUMBER_OF_RANKS, RW_MGR_MEM_NUMBER_OF_CS_PER_DIMM,
3771 RW_MGR_MEM_DQ_PER_READ_DQS, RW_MGR_MEM_DQ_PER_WRITE_DQS,
3772 RW_MGR_MEM_VIRTUAL_GROUPS_PER_READ_DQS,
3773 RW_MGR_MEM_VIRTUAL_GROUPS_PER_WRITE_DQS);
3774 debug_cond(DLEVEL == 1,
3775 "dqs=%u,%u dq=%u dm=%u ptap_delay=%u dtap_delay=%u ",
3776 RW_MGR_MEM_IF_READ_DQS_WIDTH, RW_MGR_MEM_IF_WRITE_DQS_WIDTH,
3777 RW_MGR_MEM_DATA_WIDTH, RW_MGR_MEM_DATA_MASK_WIDTH,
3778 IO_DELAY_PER_OPA_TAP, IO_DELAY_PER_DCHAIN_TAP);
3779 debug_cond(DLEVEL == 1, "dtap_dqsen_delay=%u, dll=%u",
3780 IO_DELAY_PER_DQS_EN_DCHAIN_TAP, IO_DLL_CHAIN_LENGTH);
3781 debug_cond(DLEVEL == 1, "max values: en_p=%u dqdqs_p=%u en_d=%u dqs_in_d=%u ",
3782 IO_DQS_EN_PHASE_MAX, IO_DQDQS_OUT_PHASE_MAX,
3783 IO_DQS_EN_DELAY_MAX, IO_DQS_IN_DELAY_MAX);
3784 debug_cond(DLEVEL == 1, "io_in_d=%u io_out1_d=%u io_out2_d=%u ",
3785 IO_IO_IN_DELAY_MAX, IO_IO_OUT1_DELAY_MAX,
3786 IO_IO_OUT2_DELAY_MAX);
3787 debug_cond(DLEVEL == 1, "dqs_in_reserve=%u dqs_out_reserve=%u\n",
3788 IO_DQS_IN_RESERVE, IO_DQS_OUT_RESERVE);
3790 hc_initialize_rom_data();
3792 /* update info for sims */
3793 reg_file_set_stage(CAL_STAGE_NIL);
3794 reg_file_set_group(0);
3797 * Load global needed for those actions that require
3798 * some dynamic calibration support.
3800 dyn_calib_steps = STATIC_CALIB_STEPS;
3802 * Load global to allow dynamic selection of delay loop settings
3803 * based on calibration mode.
3805 if (!(dyn_calib_steps & CALIB_SKIP_DELAY_LOOPS))
3806 skip_delay_mask = 0xff;
3808 skip_delay_mask = 0x0;
3810 pass = run_mem_calibrate();
3811 debug_mem_calibrate(pass);