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__);
1335 static int sdr_working_phase(uint32_t grp, uint32_t *bit_chk,
1336 uint32_t dtaps_per_ptap, uint32_t *work_bgn,
1337 uint32_t *v, uint32_t *d, uint32_t *p,
1338 uint32_t *i, uint32_t *max_working_cnt)
1340 uint32_t found_begin = 0;
1341 uint32_t tmp_delay = 0;
1342 uint32_t test_status;
1344 for (*d = 0; *d <= dtaps_per_ptap; (*d)++, tmp_delay +=
1345 IO_DELAY_PER_DQS_EN_DCHAIN_TAP) {
1346 *work_bgn = tmp_delay;
1347 scc_mgr_set_dqs_en_delay_all_ranks(grp, *d);
1349 for (*i = 0; *i < VFIFO_SIZE; (*i)++) {
1350 for (*p = 0; *p <= IO_DQS_EN_PHASE_MAX; (*p)++, *work_bgn +=
1351 IO_DELAY_PER_OPA_TAP) {
1352 scc_mgr_set_dqs_en_phase_all_ranks(grp, *p);
1355 rw_mgr_mem_calibrate_read_test_all_ranks
1356 (grp, 1, PASS_ONE_BIT, bit_chk, 0);
1359 *max_working_cnt = 1;
1368 if (*p > IO_DQS_EN_PHASE_MAX)
1369 /* fiddle with FIFO */
1370 rw_mgr_incr_vfifo(grp, v);
1377 if (*i >= VFIFO_SIZE) {
1378 /* cannot find working solution */
1379 debug_cond(DLEVEL == 2, "%s:%d find_dqs_en_phase: no vfifo/\
1380 ptap/dtap\n", __func__, __LINE__);
1387 static void sdr_backup_phase(uint32_t grp, uint32_t *bit_chk,
1388 uint32_t *work_bgn, uint32_t *v, uint32_t *d,
1389 uint32_t *p, uint32_t *max_working_cnt)
1391 uint32_t found_begin = 0;
1394 /* Special case code for backing up a phase */
1396 *p = IO_DQS_EN_PHASE_MAX;
1397 rw_mgr_decr_vfifo(grp, v);
1401 tmp_delay = *work_bgn - IO_DELAY_PER_OPA_TAP;
1402 scc_mgr_set_dqs_en_phase_all_ranks(grp, *p);
1404 for (*d = 0; *d <= IO_DQS_EN_DELAY_MAX && tmp_delay < *work_bgn;
1405 (*d)++, tmp_delay += IO_DELAY_PER_DQS_EN_DCHAIN_TAP) {
1406 scc_mgr_set_dqs_en_delay_all_ranks(grp, *d);
1408 if (rw_mgr_mem_calibrate_read_test_all_ranks(grp, 1,
1412 *work_bgn = tmp_delay;
1417 /* We have found a working dtap before the ptap found above */
1418 if (found_begin == 1)
1419 (*max_working_cnt)++;
1422 * Restore VFIFO to old state before we decremented it
1426 if (*p > IO_DQS_EN_PHASE_MAX) {
1428 rw_mgr_incr_vfifo(grp, v);
1431 scc_mgr_set_dqs_en_delay_all_ranks(grp, 0);
1434 static int sdr_nonworking_phase(uint32_t grp, uint32_t *bit_chk,
1435 uint32_t *work_bgn, uint32_t *v, uint32_t *d,
1436 uint32_t *p, uint32_t *i, uint32_t *max_working_cnt,
1439 uint32_t found_end = 0;
1442 *work_end += IO_DELAY_PER_OPA_TAP;
1443 if (*p > IO_DQS_EN_PHASE_MAX) {
1444 /* fiddle with FIFO */
1446 rw_mgr_incr_vfifo(grp, v);
1449 for (; *i < VFIFO_SIZE + 1; (*i)++) {
1450 for (; *p <= IO_DQS_EN_PHASE_MAX; (*p)++, *work_end
1451 += IO_DELAY_PER_OPA_TAP) {
1452 scc_mgr_set_dqs_en_phase_all_ranks(grp, *p);
1454 if (!rw_mgr_mem_calibrate_read_test_all_ranks
1455 (grp, 1, PASS_ONE_BIT, bit_chk, 0)) {
1459 (*max_working_cnt)++;
1466 if (*p > IO_DQS_EN_PHASE_MAX) {
1467 /* fiddle with FIFO */
1468 rw_mgr_incr_vfifo(grp, v);
1473 if (*i >= VFIFO_SIZE + 1) {
1474 /* cannot see edge of failing read */
1475 debug_cond(DLEVEL == 2, "%s:%d sdr_nonworking_phase: end:\
1476 failed\n", __func__, __LINE__);
1484 * sdr_find_window_center() - Find center of the working DQS window.
1485 * @grp: Read/Write group
1486 * @work_bgn: First working settings
1487 * @work_end: Last working settings
1490 * Find center of the working DQS enable window.
1492 static int sdr_find_window_center(const u32 grp, const u32 work_bgn,
1493 const u32 work_end, const u32 val)
1495 u32 bit_chk, work_mid, v = val;
1499 work_mid = (work_bgn + work_end) / 2;
1501 debug_cond(DLEVEL == 2, "work_bgn=%d work_end=%d work_mid=%d\n",
1502 work_bgn, work_end, work_mid);
1503 /* Get the middle delay to be less than a VFIFO delay */
1504 tmp_delay = (IO_DQS_EN_PHASE_MAX + 1) * IO_DELAY_PER_OPA_TAP;
1506 debug_cond(DLEVEL == 2, "vfifo ptap delay %d\n", tmp_delay);
1507 work_mid %= tmp_delay;
1508 debug_cond(DLEVEL == 2, "new work_mid %d\n", work_mid);
1510 tmp_delay = rounddown(work_mid, IO_DELAY_PER_OPA_TAP);
1511 if (tmp_delay > IO_DQS_EN_PHASE_MAX * IO_DELAY_PER_OPA_TAP)
1512 tmp_delay = IO_DQS_EN_PHASE_MAX * IO_DELAY_PER_OPA_TAP;
1513 p = tmp_delay / IO_DELAY_PER_OPA_TAP;
1515 debug_cond(DLEVEL == 2, "new p %d, tmp_delay=%d\n", p, tmp_delay);
1517 d = DIV_ROUND_UP(work_mid - tmp_delay, IO_DELAY_PER_DQS_EN_DCHAIN_TAP);
1518 if (d > IO_DQS_EN_DELAY_MAX)
1519 d = IO_DQS_EN_DELAY_MAX;
1520 tmp_delay += d * IO_DELAY_PER_DQS_EN_DCHAIN_TAP;
1522 debug_cond(DLEVEL == 2, "new d %d, tmp_delay=%d\n", d, tmp_delay);
1524 scc_mgr_set_dqs_en_phase_all_ranks(grp, p);
1525 scc_mgr_set_dqs_en_delay_all_ranks(grp, d);
1528 * push vfifo until we can successfully calibrate. We can do this
1529 * because the largest possible margin in 1 VFIFO cycle.
1531 for (i = 0; i < VFIFO_SIZE; i++) {
1532 debug_cond(DLEVEL == 2, "find_dqs_en_phase: center: vfifo=%u\n",
1534 if (rw_mgr_mem_calibrate_read_test_all_ranks(grp, 1,
1537 debug_cond(DLEVEL == 2,
1538 "%s:%d center: found: vfifo=%u ptap=%u dtap=%u\n",
1539 __func__, __LINE__, v, p, d);
1543 /* Fiddle with FIFO. */
1544 rw_mgr_incr_vfifo(grp, &v);
1547 debug_cond(DLEVEL == 2, "%s:%d center: failed.\n",
1548 __func__, __LINE__);
1552 /* find a good dqs enable to use */
1553 static uint32_t rw_mgr_mem_calibrate_vfifo_find_dqs_en_phase(uint32_t grp)
1555 uint32_t v, d, p, i;
1556 uint32_t max_working_cnt;
1558 uint32_t dtaps_per_ptap;
1559 uint32_t work_bgn, work_end;
1560 uint32_t found_passing_read, found_failing_read, initial_failing_dtap;
1562 debug("%s:%d %u\n", __func__, __LINE__, grp);
1564 reg_file_set_sub_stage(CAL_SUBSTAGE_VFIFO_CENTER);
1566 scc_mgr_set_dqs_en_delay_all_ranks(grp, 0);
1567 scc_mgr_set_dqs_en_phase_all_ranks(grp, 0);
1569 /* ************************************************************** */
1570 /* * Step 0 : Determine number of delay taps for each phase tap * */
1571 dtaps_per_ptap = IO_DELAY_PER_OPA_TAP/IO_DELAY_PER_DQS_EN_DCHAIN_TAP;
1573 /* ********************************************************* */
1574 /* * Step 1 : First push vfifo until we get a failing read * */
1575 v = find_vfifo_read(grp, &bit_chk);
1577 max_working_cnt = 0;
1579 /* ******************************************************** */
1580 /* * step 2: find first working phase, increment in ptaps * */
1582 if (sdr_working_phase(grp, &bit_chk, dtaps_per_ptap, &work_bgn, &v, &d,
1583 &p, &i, &max_working_cnt) == 0)
1586 work_end = work_bgn;
1589 * If d is 0 then the working window covers a phase tap and
1590 * we can follow the old procedure otherwise, we've found the beginning,
1591 * and we need to increment the dtaps until we find the end.
1594 /* ********************************************************* */
1595 /* * step 3a: if we have room, back off by one and
1596 increment in dtaps * */
1598 sdr_backup_phase(grp, &bit_chk, &work_bgn, &v, &d, &p,
1601 /* ********************************************************* */
1602 /* * step 4a: go forward from working phase to non working
1603 phase, increment in ptaps * */
1604 if (sdr_nonworking_phase(grp, &bit_chk, &work_bgn, &v, &d, &p,
1605 &i, &max_working_cnt, &work_end) == 0)
1608 /* ********************************************************* */
1609 /* * step 5a: back off one from last, increment in dtaps * */
1611 /* Special case code for backing up a phase */
1613 p = IO_DQS_EN_PHASE_MAX;
1614 rw_mgr_decr_vfifo(grp, &v);
1619 work_end -= IO_DELAY_PER_OPA_TAP;
1620 scc_mgr_set_dqs_en_phase_all_ranks(grp, p);
1622 /* * The actual increment of dtaps is done outside of
1623 the if/else loop to share code */
1626 debug_cond(DLEVEL == 2, "%s:%d find_dqs_en_phase: v/p: \
1627 vfifo=%u ptap=%u\n", __func__, __LINE__,
1630 /* ******************************************************* */
1631 /* * step 3-5b: Find the right edge of the window using
1633 debug_cond(DLEVEL == 2, "%s:%d find_dqs_en_phase:vfifo=%u \
1634 ptap=%u dtap=%u bgn=%u\n", __func__, __LINE__,
1637 work_end = work_bgn;
1639 /* * The actual increment of dtaps is done outside of the
1640 if/else loop to share code */
1642 /* Only here to counterbalance a subtract later on which is
1643 not needed if this branch of the algorithm is taken */
1647 /* The dtap increment to find the failing edge is done here */
1648 for (; d <= IO_DQS_EN_DELAY_MAX; d++, work_end +=
1649 IO_DELAY_PER_DQS_EN_DCHAIN_TAP) {
1650 debug_cond(DLEVEL == 2, "%s:%d find_dqs_en_phase: \
1651 end-2: dtap=%u\n", __func__, __LINE__, d);
1652 scc_mgr_set_dqs_en_delay_all_ranks(grp, d);
1654 if (!rw_mgr_mem_calibrate_read_test_all_ranks(grp, 1,
1661 /* Go back to working dtap */
1663 work_end -= IO_DELAY_PER_DQS_EN_DCHAIN_TAP;
1665 debug_cond(DLEVEL == 2, "%s:%d find_dqs_en_phase: v/p/d: vfifo=%u \
1666 ptap=%u dtap=%u end=%u\n", __func__, __LINE__,
1667 v, p, d-1, work_end);
1669 if (work_end < work_bgn) {
1671 debug_cond(DLEVEL == 2, "%s:%d find_dqs_en_phase: end-2: \
1672 failed\n", __func__, __LINE__);
1676 debug_cond(DLEVEL == 2, "%s:%d find_dqs_en_phase: found range [%u,%u]\n",
1677 __func__, __LINE__, work_bgn, work_end);
1679 /* *************************************************************** */
1681 * * We need to calculate the number of dtaps that equal a ptap
1682 * * To do that we'll back up a ptap and re-find the edge of the
1683 * * window using dtaps
1686 debug_cond(DLEVEL == 2, "%s:%d find_dqs_en_phase: calculate dtaps_per_ptap \
1687 for tracking\n", __func__, __LINE__);
1689 /* Special case code for backing up a phase */
1691 p = IO_DQS_EN_PHASE_MAX;
1692 rw_mgr_decr_vfifo(grp, &v);
1693 debug_cond(DLEVEL == 2, "%s:%d find_dqs_en_phase: backedup \
1694 cycle/phase: v=%u p=%u\n", __func__, __LINE__,
1698 debug_cond(DLEVEL == 2, "%s:%d find_dqs_en_phase: backedup \
1699 phase only: v=%u p=%u", __func__, __LINE__,
1703 scc_mgr_set_dqs_en_phase_all_ranks(grp, p);
1706 * Increase dtap until we first see a passing read (in case the
1707 * window is smaller than a ptap),
1708 * and then a failing read to mark the edge of the window again
1711 /* Find a passing read */
1712 debug_cond(DLEVEL == 2, "%s:%d find_dqs_en_phase: find passing read\n",
1713 __func__, __LINE__);
1714 found_passing_read = 0;
1715 found_failing_read = 0;
1716 initial_failing_dtap = d;
1717 for (; d <= IO_DQS_EN_DELAY_MAX; d++) {
1718 debug_cond(DLEVEL == 2, "%s:%d find_dqs_en_phase: testing \
1719 read d=%u\n", __func__, __LINE__, d);
1720 scc_mgr_set_dqs_en_delay_all_ranks(grp, d);
1722 if (rw_mgr_mem_calibrate_read_test_all_ranks(grp, 1,
1725 found_passing_read = 1;
1730 if (found_passing_read) {
1731 /* Find a failing read */
1732 debug_cond(DLEVEL == 2, "%s:%d find_dqs_en_phase: find failing \
1733 read\n", __func__, __LINE__);
1734 for (d = d + 1; d <= IO_DQS_EN_DELAY_MAX; d++) {
1735 debug_cond(DLEVEL == 2, "%s:%d find_dqs_en_phase: \
1736 testing read d=%u\n", __func__, __LINE__, d);
1737 scc_mgr_set_dqs_en_delay_all_ranks(grp, d);
1739 if (!rw_mgr_mem_calibrate_read_test_all_ranks
1740 (grp, 1, PASS_ONE_BIT, &bit_chk, 0)) {
1741 found_failing_read = 1;
1746 debug_cond(DLEVEL == 1, "%s:%d find_dqs_en_phase: failed to \
1747 calculate dtaps", __func__, __LINE__);
1748 debug_cond(DLEVEL == 1, "per ptap. Fall back on static value\n");
1752 * The dynamically calculated dtaps_per_ptap is only valid if we
1753 * found a passing/failing read. If we didn't, it means d hit the max
1754 * (IO_DQS_EN_DELAY_MAX). Otherwise, dtaps_per_ptap retains its
1755 * statically calculated value.
1757 if (found_passing_read && found_failing_read)
1758 dtaps_per_ptap = d - initial_failing_dtap;
1760 writel(dtaps_per_ptap, &sdr_reg_file->dtaps_per_ptap);
1761 debug_cond(DLEVEL == 2, "%s:%d find_dqs_en_phase: dtaps_per_ptap=%u \
1762 - %u = %u", __func__, __LINE__, d,
1763 initial_failing_dtap, dtaps_per_ptap);
1765 /* ******************************************** */
1766 /* * step 6: Find the centre of the window * */
1767 if (sdr_find_window_centre(grp, work_bgn, work_end, v))
1768 return 0; /* FIXME: Old code, return 0 means failure :-( */
1773 /* per-bit deskew DQ and center */
1774 static uint32_t rw_mgr_mem_calibrate_vfifo_center(uint32_t rank_bgn,
1775 uint32_t write_group, uint32_t read_group, uint32_t test_bgn,
1776 uint32_t use_read_test, uint32_t update_fom)
1778 uint32_t i, p, d, min_index;
1780 * Store these as signed since there are comparisons with
1784 uint32_t sticky_bit_chk;
1785 int32_t left_edge[RW_MGR_MEM_DQ_PER_READ_DQS];
1786 int32_t right_edge[RW_MGR_MEM_DQ_PER_READ_DQS];
1787 int32_t final_dq[RW_MGR_MEM_DQ_PER_READ_DQS];
1789 int32_t orig_mid_min, mid_min;
1790 int32_t new_dqs, start_dqs, start_dqs_en, shift_dq, final_dqs,
1792 int32_t dq_margin, dqs_margin;
1794 uint32_t temp_dq_in_delay1, temp_dq_in_delay2;
1797 debug("%s:%d: %u %u", __func__, __LINE__, read_group, test_bgn);
1799 addr = SDR_PHYGRP_SCCGRP_ADDRESS | SCC_MGR_DQS_IN_DELAY_OFFSET;
1800 start_dqs = readl(addr + (read_group << 2));
1801 if (IO_SHIFT_DQS_EN_WHEN_SHIFT_DQS)
1802 start_dqs_en = readl(addr + ((read_group << 2)
1803 - IO_DQS_EN_DELAY_OFFSET));
1805 /* set the left and right edge of each bit to an illegal value */
1806 /* use (IO_IO_IN_DELAY_MAX + 1) as an illegal value */
1808 for (i = 0; i < RW_MGR_MEM_DQ_PER_READ_DQS; i++) {
1809 left_edge[i] = IO_IO_IN_DELAY_MAX + 1;
1810 right_edge[i] = IO_IO_IN_DELAY_MAX + 1;
1813 /* Search for the left edge of the window for each bit */
1814 for (d = 0; d <= IO_IO_IN_DELAY_MAX; d++) {
1815 scc_mgr_apply_group_dq_in_delay(write_group, test_bgn, d);
1817 writel(0, &sdr_scc_mgr->update);
1820 * Stop searching when the read test doesn't pass AND when
1821 * we've seen a passing read on every bit.
1823 if (use_read_test) {
1824 stop = !rw_mgr_mem_calibrate_read_test(rank_bgn,
1825 read_group, NUM_READ_PB_TESTS, PASS_ONE_BIT,
1828 rw_mgr_mem_calibrate_write_test(rank_bgn, write_group,
1831 bit_chk = bit_chk >> (RW_MGR_MEM_DQ_PER_READ_DQS *
1832 (read_group - (write_group *
1833 RW_MGR_MEM_IF_READ_DQS_WIDTH /
1834 RW_MGR_MEM_IF_WRITE_DQS_WIDTH)));
1835 stop = (bit_chk == 0);
1837 sticky_bit_chk = sticky_bit_chk | bit_chk;
1838 stop = stop && (sticky_bit_chk == param->read_correct_mask);
1839 debug_cond(DLEVEL == 2, "%s:%d vfifo_center(left): dtap=%u => %u == %u \
1840 && %u", __func__, __LINE__, d,
1842 param->read_correct_mask, stop);
1847 for (i = 0; i < RW_MGR_MEM_DQ_PER_READ_DQS; i++) {
1849 /* Remember a passing test as the
1853 /* If a left edge has not been seen yet,
1854 then a future passing test will mark
1855 this edge as the right edge */
1857 IO_IO_IN_DELAY_MAX + 1) {
1858 right_edge[i] = -(d + 1);
1861 bit_chk = bit_chk >> 1;
1866 /* Reset DQ delay chains to 0 */
1867 scc_mgr_apply_group_dq_in_delay(test_bgn, 0);
1869 for (i = RW_MGR_MEM_DQ_PER_READ_DQS - 1;; i--) {
1870 debug_cond(DLEVEL == 2, "%s:%d vfifo_center: left_edge[%u]: \
1871 %d right_edge[%u]: %d\n", __func__, __LINE__,
1872 i, left_edge[i], i, right_edge[i]);
1875 * Check for cases where we haven't found the left edge,
1876 * which makes our assignment of the the right edge invalid.
1877 * Reset it to the illegal value.
1879 if ((left_edge[i] == IO_IO_IN_DELAY_MAX + 1) && (
1880 right_edge[i] != IO_IO_IN_DELAY_MAX + 1)) {
1881 right_edge[i] = IO_IO_IN_DELAY_MAX + 1;
1882 debug_cond(DLEVEL == 2, "%s:%d vfifo_center: reset \
1883 right_edge[%u]: %d\n", __func__, __LINE__,
1888 * Reset sticky bit (except for bits where we have seen
1889 * both the left and right edge).
1891 sticky_bit_chk = sticky_bit_chk << 1;
1892 if ((left_edge[i] != IO_IO_IN_DELAY_MAX + 1) &&
1893 (right_edge[i] != IO_IO_IN_DELAY_MAX + 1)) {
1894 sticky_bit_chk = sticky_bit_chk | 1;
1901 /* Search for the right edge of the window for each bit */
1902 for (d = 0; d <= IO_DQS_IN_DELAY_MAX - start_dqs; d++) {
1903 scc_mgr_set_dqs_bus_in_delay(read_group, d + start_dqs);
1904 if (IO_SHIFT_DQS_EN_WHEN_SHIFT_DQS) {
1905 uint32_t delay = d + start_dqs_en;
1906 if (delay > IO_DQS_EN_DELAY_MAX)
1907 delay = IO_DQS_EN_DELAY_MAX;
1908 scc_mgr_set_dqs_en_delay(read_group, delay);
1910 scc_mgr_load_dqs(read_group);
1912 writel(0, &sdr_scc_mgr->update);
1915 * Stop searching when the read test doesn't pass AND when
1916 * we've seen a passing read on every bit.
1918 if (use_read_test) {
1919 stop = !rw_mgr_mem_calibrate_read_test(rank_bgn,
1920 read_group, NUM_READ_PB_TESTS, PASS_ONE_BIT,
1923 rw_mgr_mem_calibrate_write_test(rank_bgn, write_group,
1926 bit_chk = bit_chk >> (RW_MGR_MEM_DQ_PER_READ_DQS *
1927 (read_group - (write_group *
1928 RW_MGR_MEM_IF_READ_DQS_WIDTH /
1929 RW_MGR_MEM_IF_WRITE_DQS_WIDTH)));
1930 stop = (bit_chk == 0);
1932 sticky_bit_chk = sticky_bit_chk | bit_chk;
1933 stop = stop && (sticky_bit_chk == param->read_correct_mask);
1935 debug_cond(DLEVEL == 2, "%s:%d vfifo_center(right): dtap=%u => %u == \
1936 %u && %u", __func__, __LINE__, d,
1937 sticky_bit_chk, param->read_correct_mask, stop);
1942 for (i = 0; i < RW_MGR_MEM_DQ_PER_READ_DQS; i++) {
1944 /* Remember a passing test as
1949 /* If a right edge has not been
1950 seen yet, then a future passing
1951 test will mark this edge as the
1953 if (right_edge[i] ==
1954 IO_IO_IN_DELAY_MAX + 1) {
1955 left_edge[i] = -(d + 1);
1958 /* d = 0 failed, but it passed
1959 when testing the left edge,
1960 so it must be marginal,
1962 if (right_edge[i] ==
1963 IO_IO_IN_DELAY_MAX + 1 &&
1969 /* If a right edge has not been
1970 seen yet, then a future passing
1971 test will mark this edge as the
1973 else if (right_edge[i] ==
1974 IO_IO_IN_DELAY_MAX +
1976 left_edge[i] = -(d + 1);
1981 debug_cond(DLEVEL == 2, "%s:%d vfifo_center[r,\
1982 d=%u]: ", __func__, __LINE__, d);
1983 debug_cond(DLEVEL == 2, "bit_chk_test=%d left_edge[%u]: %d ",
1984 (int)(bit_chk & 1), i, left_edge[i]);
1985 debug_cond(DLEVEL == 2, "right_edge[%u]: %d\n", i,
1987 bit_chk = bit_chk >> 1;
1992 /* Check that all bits have a window */
1993 for (i = 0; i < RW_MGR_MEM_DQ_PER_READ_DQS; i++) {
1994 debug_cond(DLEVEL == 2, "%s:%d vfifo_center: left_edge[%u]: \
1995 %d right_edge[%u]: %d", __func__, __LINE__,
1996 i, left_edge[i], i, right_edge[i]);
1997 if ((left_edge[i] == IO_IO_IN_DELAY_MAX + 1) || (right_edge[i]
1998 == IO_IO_IN_DELAY_MAX + 1)) {
2000 * Restore delay chain settings before letting the loop
2001 * in rw_mgr_mem_calibrate_vfifo to retry different
2002 * dqs/ck relationships.
2004 scc_mgr_set_dqs_bus_in_delay(read_group, start_dqs);
2005 if (IO_SHIFT_DQS_EN_WHEN_SHIFT_DQS) {
2006 scc_mgr_set_dqs_en_delay(read_group,
2009 scc_mgr_load_dqs(read_group);
2010 writel(0, &sdr_scc_mgr->update);
2012 debug_cond(DLEVEL == 1, "%s:%d vfifo_center: failed to \
2013 find edge [%u]: %d %d", __func__, __LINE__,
2014 i, left_edge[i], right_edge[i]);
2015 if (use_read_test) {
2016 set_failing_group_stage(read_group *
2017 RW_MGR_MEM_DQ_PER_READ_DQS + i,
2019 CAL_SUBSTAGE_VFIFO_CENTER);
2021 set_failing_group_stage(read_group *
2022 RW_MGR_MEM_DQ_PER_READ_DQS + i,
2023 CAL_STAGE_VFIFO_AFTER_WRITES,
2024 CAL_SUBSTAGE_VFIFO_CENTER);
2030 /* Find middle of window for each DQ bit */
2031 mid_min = left_edge[0] - right_edge[0];
2033 for (i = 1; i < RW_MGR_MEM_DQ_PER_READ_DQS; i++) {
2034 mid = left_edge[i] - right_edge[i];
2035 if (mid < mid_min) {
2042 * -mid_min/2 represents the amount that we need to move DQS.
2043 * If mid_min is odd and positive we'll need to add one to
2044 * make sure the rounding in further calculations is correct
2045 * (always bias to the right), so just add 1 for all positive values.
2050 mid_min = mid_min / 2;
2052 debug_cond(DLEVEL == 1, "%s:%d vfifo_center: mid_min=%d (index=%u)\n",
2053 __func__, __LINE__, mid_min, min_index);
2055 /* Determine the amount we can change DQS (which is -mid_min) */
2056 orig_mid_min = mid_min;
2057 new_dqs = start_dqs - mid_min;
2058 if (new_dqs > IO_DQS_IN_DELAY_MAX)
2059 new_dqs = IO_DQS_IN_DELAY_MAX;
2060 else if (new_dqs < 0)
2063 mid_min = start_dqs - new_dqs;
2064 debug_cond(DLEVEL == 1, "vfifo_center: new mid_min=%d new_dqs=%d\n",
2067 if (IO_SHIFT_DQS_EN_WHEN_SHIFT_DQS) {
2068 if (start_dqs_en - mid_min > IO_DQS_EN_DELAY_MAX)
2069 mid_min += start_dqs_en - mid_min - IO_DQS_EN_DELAY_MAX;
2070 else if (start_dqs_en - mid_min < 0)
2071 mid_min += start_dqs_en - mid_min;
2073 new_dqs = start_dqs - mid_min;
2075 debug_cond(DLEVEL == 1, "vfifo_center: start_dqs=%d start_dqs_en=%d \
2076 new_dqs=%d mid_min=%d\n", start_dqs,
2077 IO_SHIFT_DQS_EN_WHEN_SHIFT_DQS ? start_dqs_en : -1,
2080 /* Initialize data for export structures */
2081 dqs_margin = IO_IO_IN_DELAY_MAX + 1;
2082 dq_margin = IO_IO_IN_DELAY_MAX + 1;
2084 /* add delay to bring centre of all DQ windows to the same "level" */
2085 for (i = 0, p = test_bgn; i < RW_MGR_MEM_DQ_PER_READ_DQS; i++, p++) {
2086 /* Use values before divide by 2 to reduce round off error */
2087 shift_dq = (left_edge[i] - right_edge[i] -
2088 (left_edge[min_index] - right_edge[min_index]))/2 +
2089 (orig_mid_min - mid_min);
2091 debug_cond(DLEVEL == 2, "vfifo_center: before: \
2092 shift_dq[%u]=%d\n", i, shift_dq);
2094 addr = SDR_PHYGRP_SCCGRP_ADDRESS | SCC_MGR_IO_IN_DELAY_OFFSET;
2095 temp_dq_in_delay1 = readl(addr + (p << 2));
2096 temp_dq_in_delay2 = readl(addr + (i << 2));
2098 if (shift_dq + (int32_t)temp_dq_in_delay1 >
2099 (int32_t)IO_IO_IN_DELAY_MAX) {
2100 shift_dq = (int32_t)IO_IO_IN_DELAY_MAX - temp_dq_in_delay2;
2101 } else if (shift_dq + (int32_t)temp_dq_in_delay1 < 0) {
2102 shift_dq = -(int32_t)temp_dq_in_delay1;
2104 debug_cond(DLEVEL == 2, "vfifo_center: after: \
2105 shift_dq[%u]=%d\n", i, shift_dq);
2106 final_dq[i] = temp_dq_in_delay1 + shift_dq;
2107 scc_mgr_set_dq_in_delay(p, final_dq[i]);
2110 debug_cond(DLEVEL == 2, "vfifo_center: margin[%u]=[%d,%d]\n", i,
2111 left_edge[i] - shift_dq + (-mid_min),
2112 right_edge[i] + shift_dq - (-mid_min));
2113 /* To determine values for export structures */
2114 if (left_edge[i] - shift_dq + (-mid_min) < dq_margin)
2115 dq_margin = left_edge[i] - shift_dq + (-mid_min);
2117 if (right_edge[i] + shift_dq - (-mid_min) < dqs_margin)
2118 dqs_margin = right_edge[i] + shift_dq - (-mid_min);
2121 final_dqs = new_dqs;
2122 if (IO_SHIFT_DQS_EN_WHEN_SHIFT_DQS)
2123 final_dqs_en = start_dqs_en - mid_min;
2126 if (IO_SHIFT_DQS_EN_WHEN_SHIFT_DQS) {
2127 scc_mgr_set_dqs_en_delay(read_group, final_dqs_en);
2128 scc_mgr_load_dqs(read_group);
2132 scc_mgr_set_dqs_bus_in_delay(read_group, final_dqs);
2133 scc_mgr_load_dqs(read_group);
2134 debug_cond(DLEVEL == 2, "%s:%d vfifo_center: dq_margin=%d \
2135 dqs_margin=%d", __func__, __LINE__,
2136 dq_margin, dqs_margin);
2139 * Do not remove this line as it makes sure all of our decisions
2140 * have been applied. Apply the update bit.
2142 writel(0, &sdr_scc_mgr->update);
2144 return (dq_margin >= 0) && (dqs_margin >= 0);
2148 * rw_mgr_mem_calibrate_guaranteed_write() - Perform guaranteed write into the device
2149 * @rw_group: Read/Write Group
2150 * @phase: DQ/DQS phase
2152 * Because initially no communication ca be reliably performed with the memory
2153 * device, the sequencer uses a guaranteed write mechanism to write data into
2154 * the memory device.
2156 static int rw_mgr_mem_calibrate_guaranteed_write(const u32 rw_group,
2161 /* Set a particular DQ/DQS phase. */
2162 scc_mgr_set_dqdqs_output_phase_all_ranks(rw_group, phase);
2164 debug_cond(DLEVEL == 1, "%s:%d guaranteed write: g=%u p=%u\n",
2165 __func__, __LINE__, rw_group, phase);
2168 * Altera EMI_RM 2015.05.04 :: Figure 1-25
2169 * Load up the patterns used by read calibration using the
2170 * current DQDQS phase.
2172 rw_mgr_mem_calibrate_read_load_patterns(0, 1);
2174 if (gbl->phy_debug_mode_flags & PHY_DEBUG_DISABLE_GUARANTEED_READ)
2178 * Altera EMI_RM 2015.05.04 :: Figure 1-26
2179 * Back-to-Back reads of the patterns used for calibration.
2181 ret = rw_mgr_mem_calibrate_read_test_patterns(0, rw_group, 1);
2183 debug_cond(DLEVEL == 1,
2184 "%s:%d Guaranteed read test failed: g=%u p=%u\n",
2185 __func__, __LINE__, rw_group, phase);
2190 * rw_mgr_mem_calibrate_dqs_enable_calibration() - DQS Enable Calibration
2191 * @rw_group: Read/Write Group
2192 * @test_bgn: Rank at which the test begins
2194 * DQS enable calibration ensures reliable capture of the DQ signal without
2195 * glitches on the DQS line.
2197 static int rw_mgr_mem_calibrate_dqs_enable_calibration(const u32 rw_group,
2201 * Altera EMI_RM 2015.05.04 :: Figure 1-27
2202 * DQS and DQS Eanble Signal Relationships.
2205 /* We start at zero, so have one less dq to devide among */
2206 const u32 delay_step = IO_IO_IN_DELAY_MAX /
2207 (RW_MGR_MEM_DQ_PER_READ_DQS - 1);
2211 debug("%s:%d (%u,%u)\n", __func__, __LINE__, rw_group, test_bgn);
2213 /* Try different dq_in_delays since the DQ path is shorter than DQS. */
2214 for (r = 0; r < RW_MGR_MEM_NUMBER_OF_RANKS;
2215 r += NUM_RANKS_PER_SHADOW_REG) {
2216 for (i = 0, p = test_bgn, d = 0;
2217 i < RW_MGR_MEM_DQ_PER_READ_DQS;
2218 i++, p++, d += delay_step) {
2219 debug_cond(DLEVEL == 1,
2220 "%s:%d: g=%u r=%u i=%u p=%u d=%u\n",
2221 __func__, __LINE__, rw_group, r, i, p, d);
2223 scc_mgr_set_dq_in_delay(p, d);
2227 writel(0, &sdr_scc_mgr->update);
2231 * Try rw_mgr_mem_calibrate_vfifo_find_dqs_en_phase across different
2232 * dq_in_delay values
2234 found = rw_mgr_mem_calibrate_vfifo_find_dqs_en_phase(rw_group);
2236 debug_cond(DLEVEL == 1,
2237 "%s:%d: g=%u found=%u; Reseting delay chain to zero\n",
2238 __func__, __LINE__, rw_group, found);
2240 for (r = 0; r < RW_MGR_MEM_NUMBER_OF_RANKS;
2241 r += NUM_RANKS_PER_SHADOW_REG) {
2242 scc_mgr_apply_group_dq_in_delay(test_bgn, 0);
2243 writel(0, &sdr_scc_mgr->update);
2254 * rw_mgr_mem_calibrate_dq_dqs_centering() - Centering DQ/DQS
2255 * @rw_group: Read/Write Group
2256 * @test_bgn: Rank at which the test begins
2257 * @use_read_test: Perform a read test
2258 * @update_fom: Update FOM
2260 * The centerin DQ/DQS stage attempts to align DQ and DQS signals on reads
2264 rw_mgr_mem_calibrate_dq_dqs_centering(const u32 rw_group, const u32 test_bgn,
2265 const int use_read_test,
2266 const int update_fom)
2269 int ret, grp_calibrated;
2273 * Altera EMI_RM 2015.05.04 :: Figure 1-28
2274 * Read per-bit deskew can be done on a per shadow register basis.
2277 for (rank_bgn = 0, sr = 0;
2278 rank_bgn < RW_MGR_MEM_NUMBER_OF_RANKS;
2279 rank_bgn += NUM_RANKS_PER_SHADOW_REG, sr++) {
2280 /* Check if this set of ranks should be skipped entirely. */
2281 if (param->skip_shadow_regs[sr])
2284 ret = rw_mgr_mem_calibrate_vfifo_center(rank_bgn, rw_group,
2294 if (!grp_calibrated)
2301 * rw_mgr_mem_calibrate_vfifo() - Calibrate the read valid prediction FIFO
2302 * @rw_group: Read/Write Group
2303 * @test_bgn: Rank at which the test begins
2305 * Stage 1: Calibrate the read valid prediction FIFO.
2307 * This function implements UniPHY calibration Stage 1, as explained in
2308 * detail in Altera EMI_RM 2015.05.04 , "UniPHY Calibration Stages".
2310 * - read valid prediction will consist of finding:
2311 * - DQS enable phase and DQS enable delay (DQS Enable Calibration)
2312 * - DQS input phase and DQS input delay (DQ/DQS Centering)
2313 * - we also do a per-bit deskew on the DQ lines.
2315 static int rw_mgr_mem_calibrate_vfifo(const u32 rw_group, const u32 test_bgn)
2318 uint32_t dtaps_per_ptap;
2319 uint32_t failed_substage;
2323 debug("%s:%d: %u %u\n", __func__, __LINE__, rw_group, test_bgn);
2325 /* Update info for sims */
2326 reg_file_set_group(rw_group);
2327 reg_file_set_stage(CAL_STAGE_VFIFO);
2328 reg_file_set_sub_stage(CAL_SUBSTAGE_GUARANTEED_READ);
2330 failed_substage = CAL_SUBSTAGE_GUARANTEED_READ;
2332 /* USER Determine number of delay taps for each phase tap. */
2333 dtaps_per_ptap = DIV_ROUND_UP(IO_DELAY_PER_OPA_TAP,
2334 IO_DELAY_PER_DQS_EN_DCHAIN_TAP) - 1;
2336 for (d = 0; d <= dtaps_per_ptap; d += 2) {
2338 * In RLDRAMX we may be messing the delay of pins in
2339 * the same write rw_group but outside of the current read
2340 * the rw_group, but that's ok because we haven't calibrated
2344 scc_mgr_apply_group_all_out_delay_add_all_ranks(
2348 for (p = 0; p <= IO_DQDQS_OUT_PHASE_MAX; p++) {
2349 /* 1) Guaranteed Write */
2350 ret = rw_mgr_mem_calibrate_guaranteed_write(rw_group, p);
2354 /* 2) DQS Enable Calibration */
2355 ret = rw_mgr_mem_calibrate_dqs_enable_calibration(rw_group,
2358 failed_substage = CAL_SUBSTAGE_DQS_EN_PHASE;
2362 /* 3) Centering DQ/DQS */
2364 * If doing read after write calibration, do not update
2365 * FOM now. Do it then.
2367 ret = rw_mgr_mem_calibrate_dq_dqs_centering(rw_group,
2370 failed_substage = CAL_SUBSTAGE_VFIFO_CENTER;
2379 /* Calibration Stage 1 failed. */
2380 set_failing_group_stage(rw_group, CAL_STAGE_VFIFO, failed_substage);
2383 /* Calibration Stage 1 completed OK. */
2386 * Reset the delay chains back to zero if they have moved > 1
2387 * (check for > 1 because loop will increase d even when pass in
2391 scc_mgr_zero_group(rw_group, 1);
2396 /* VFIFO Calibration -- Read Deskew Calibration after write deskew */
2397 static uint32_t rw_mgr_mem_calibrate_vfifo_end(uint32_t read_group,
2400 uint32_t rank_bgn, sr;
2401 uint32_t grp_calibrated;
2402 uint32_t write_group;
2404 debug("%s:%d %u %u", __func__, __LINE__, read_group, test_bgn);
2406 /* update info for sims */
2408 reg_file_set_stage(CAL_STAGE_VFIFO_AFTER_WRITES);
2409 reg_file_set_sub_stage(CAL_SUBSTAGE_VFIFO_CENTER);
2411 write_group = read_group;
2413 /* update info for sims */
2414 reg_file_set_group(read_group);
2417 /* Read per-bit deskew can be done on a per shadow register basis */
2418 for (rank_bgn = 0, sr = 0; rank_bgn < RW_MGR_MEM_NUMBER_OF_RANKS;
2419 rank_bgn += NUM_RANKS_PER_SHADOW_REG, ++sr) {
2420 /* Determine if this set of ranks should be skipped entirely */
2421 if (!param->skip_shadow_regs[sr]) {
2422 /* This is the last calibration round, update FOM here */
2423 if (!rw_mgr_mem_calibrate_vfifo_center(rank_bgn,
2434 if (grp_calibrated == 0) {
2435 set_failing_group_stage(write_group,
2436 CAL_STAGE_VFIFO_AFTER_WRITES,
2437 CAL_SUBSTAGE_VFIFO_CENTER);
2444 /* Calibrate LFIFO to find smallest read latency */
2445 static uint32_t rw_mgr_mem_calibrate_lfifo(void)
2450 debug("%s:%d\n", __func__, __LINE__);
2452 /* update info for sims */
2453 reg_file_set_stage(CAL_STAGE_LFIFO);
2454 reg_file_set_sub_stage(CAL_SUBSTAGE_READ_LATENCY);
2456 /* Load up the patterns used by read calibration for all ranks */
2457 rw_mgr_mem_calibrate_read_load_patterns(0, 1);
2461 writel(gbl->curr_read_lat, &phy_mgr_cfg->phy_rlat);
2462 debug_cond(DLEVEL == 2, "%s:%d lfifo: read_lat=%u",
2463 __func__, __LINE__, gbl->curr_read_lat);
2465 if (!rw_mgr_mem_calibrate_read_test_all_ranks(0,
2473 /* reduce read latency and see if things are working */
2475 gbl->curr_read_lat--;
2476 } while (gbl->curr_read_lat > 0);
2478 /* reset the fifos to get pointers to known state */
2480 writel(0, &phy_mgr_cmd->fifo_reset);
2483 /* add a fudge factor to the read latency that was determined */
2484 gbl->curr_read_lat += 2;
2485 writel(gbl->curr_read_lat, &phy_mgr_cfg->phy_rlat);
2486 debug_cond(DLEVEL == 2, "%s:%d lfifo: success: using \
2487 read_lat=%u\n", __func__, __LINE__,
2488 gbl->curr_read_lat);
2491 set_failing_group_stage(0xff, CAL_STAGE_LFIFO,
2492 CAL_SUBSTAGE_READ_LATENCY);
2494 debug_cond(DLEVEL == 2, "%s:%d lfifo: failed at initial \
2495 read_lat=%u\n", __func__, __LINE__,
2496 gbl->curr_read_lat);
2502 * issue write test command.
2503 * two variants are provided. one that just tests a write pattern and
2504 * another that tests datamask functionality.
2506 static void rw_mgr_mem_calibrate_write_test_issue(uint32_t group,
2509 uint32_t mcc_instruction;
2510 uint32_t quick_write_mode = (((STATIC_CALIB_STEPS) & CALIB_SKIP_WRITES) &&
2511 ENABLE_SUPER_QUICK_CALIBRATION);
2512 uint32_t rw_wl_nop_cycles;
2516 * Set counter and jump addresses for the right
2517 * number of NOP cycles.
2518 * The number of supported NOP cycles can range from -1 to infinity
2519 * Three different cases are handled:
2521 * 1. For a number of NOP cycles greater than 0, the RW Mgr looping
2522 * mechanism will be used to insert the right number of NOPs
2524 * 2. For a number of NOP cycles equals to 0, the micro-instruction
2525 * issuing the write command will jump straight to the
2526 * micro-instruction that turns on DQS (for DDRx), or outputs write
2527 * data (for RLD), skipping
2528 * the NOP micro-instruction all together
2530 * 3. A number of NOP cycles equal to -1 indicates that DQS must be
2531 * turned on in the same micro-instruction that issues the write
2532 * command. Then we need
2533 * to directly jump to the micro-instruction that sends out the data
2535 * NOTE: Implementing this mechanism uses 2 RW Mgr jump-counters
2536 * (2 and 3). One jump-counter (0) is used to perform multiple
2537 * write-read operations.
2538 * one counter left to issue this command in "multiple-group" mode
2541 rw_wl_nop_cycles = gbl->rw_wl_nop_cycles;
2543 if (rw_wl_nop_cycles == -1) {
2545 * CNTR 2 - We want to execute the special write operation that
2546 * turns on DQS right away and then skip directly to the
2547 * instruction that sends out the data. We set the counter to a
2548 * large number so that the jump is always taken.
2550 writel(0xFF, &sdr_rw_load_mgr_regs->load_cntr2);
2552 /* CNTR 3 - Not used */
2554 mcc_instruction = RW_MGR_LFSR_WR_RD_DM_BANK_0_WL_1;
2555 writel(RW_MGR_LFSR_WR_RD_DM_BANK_0_DATA,
2556 &sdr_rw_load_jump_mgr_regs->load_jump_add2);
2557 writel(RW_MGR_LFSR_WR_RD_DM_BANK_0_NOP,
2558 &sdr_rw_load_jump_mgr_regs->load_jump_add3);
2560 mcc_instruction = RW_MGR_LFSR_WR_RD_BANK_0_WL_1;
2561 writel(RW_MGR_LFSR_WR_RD_BANK_0_DATA,
2562 &sdr_rw_load_jump_mgr_regs->load_jump_add2);
2563 writel(RW_MGR_LFSR_WR_RD_BANK_0_NOP,
2564 &sdr_rw_load_jump_mgr_regs->load_jump_add3);
2566 } else if (rw_wl_nop_cycles == 0) {
2568 * CNTR 2 - We want to skip the NOP operation and go straight
2569 * to the DQS enable instruction. We set the counter to a large
2570 * number so that the jump is always taken.
2572 writel(0xFF, &sdr_rw_load_mgr_regs->load_cntr2);
2574 /* CNTR 3 - Not used */
2576 mcc_instruction = RW_MGR_LFSR_WR_RD_DM_BANK_0;
2577 writel(RW_MGR_LFSR_WR_RD_DM_BANK_0_DQS,
2578 &sdr_rw_load_jump_mgr_regs->load_jump_add2);
2580 mcc_instruction = RW_MGR_LFSR_WR_RD_BANK_0;
2581 writel(RW_MGR_LFSR_WR_RD_BANK_0_DQS,
2582 &sdr_rw_load_jump_mgr_regs->load_jump_add2);
2586 * CNTR 2 - In this case we want to execute the next instruction
2587 * and NOT take the jump. So we set the counter to 0. The jump
2588 * address doesn't count.
2590 writel(0x0, &sdr_rw_load_mgr_regs->load_cntr2);
2591 writel(0x0, &sdr_rw_load_jump_mgr_regs->load_jump_add2);
2594 * CNTR 3 - Set the nop counter to the number of cycles we
2595 * need to loop for, minus 1.
2597 writel(rw_wl_nop_cycles - 1, &sdr_rw_load_mgr_regs->load_cntr3);
2599 mcc_instruction = RW_MGR_LFSR_WR_RD_DM_BANK_0;
2600 writel(RW_MGR_LFSR_WR_RD_DM_BANK_0_NOP,
2601 &sdr_rw_load_jump_mgr_regs->load_jump_add3);
2603 mcc_instruction = RW_MGR_LFSR_WR_RD_BANK_0;
2604 writel(RW_MGR_LFSR_WR_RD_BANK_0_NOP,
2605 &sdr_rw_load_jump_mgr_regs->load_jump_add3);
2609 writel(0, SDR_PHYGRP_RWMGRGRP_ADDRESS |
2610 RW_MGR_RESET_READ_DATAPATH_OFFSET);
2612 if (quick_write_mode)
2613 writel(0x08, &sdr_rw_load_mgr_regs->load_cntr0);
2615 writel(0x40, &sdr_rw_load_mgr_regs->load_cntr0);
2617 writel(mcc_instruction, &sdr_rw_load_jump_mgr_regs->load_jump_add0);
2620 * CNTR 1 - This is used to ensure enough time elapses
2621 * for read data to come back.
2623 writel(0x30, &sdr_rw_load_mgr_regs->load_cntr1);
2626 writel(RW_MGR_LFSR_WR_RD_DM_BANK_0_WAIT,
2627 &sdr_rw_load_jump_mgr_regs->load_jump_add1);
2629 writel(RW_MGR_LFSR_WR_RD_BANK_0_WAIT,
2630 &sdr_rw_load_jump_mgr_regs->load_jump_add1);
2633 addr = SDR_PHYGRP_RWMGRGRP_ADDRESS | RW_MGR_RUN_SINGLE_GROUP_OFFSET;
2634 writel(mcc_instruction, addr + (group << 2));
2637 /* Test writes, can check for a single bit pass or multiple bit pass */
2638 static uint32_t rw_mgr_mem_calibrate_write_test(uint32_t rank_bgn,
2639 uint32_t write_group, uint32_t use_dm, uint32_t all_correct,
2640 uint32_t *bit_chk, uint32_t all_ranks)
2643 uint32_t correct_mask_vg;
2644 uint32_t tmp_bit_chk;
2646 uint32_t rank_end = all_ranks ? RW_MGR_MEM_NUMBER_OF_RANKS :
2647 (rank_bgn + NUM_RANKS_PER_SHADOW_REG);
2648 uint32_t addr_rw_mgr;
2649 uint32_t base_rw_mgr;
2651 *bit_chk = param->write_correct_mask;
2652 correct_mask_vg = param->write_correct_mask_vg;
2654 for (r = rank_bgn; r < rank_end; r++) {
2655 if (param->skip_ranks[r]) {
2656 /* request to skip the rank */
2661 set_rank_and_odt_mask(r, RW_MGR_ODT_MODE_READ_WRITE);
2664 addr_rw_mgr = SDR_PHYGRP_RWMGRGRP_ADDRESS;
2665 for (vg = RW_MGR_MEM_VIRTUAL_GROUPS_PER_WRITE_DQS-1; ; vg--) {
2666 /* reset the fifos to get pointers to known state */
2667 writel(0, &phy_mgr_cmd->fifo_reset);
2669 tmp_bit_chk = tmp_bit_chk <<
2670 (RW_MGR_MEM_DQ_PER_WRITE_DQS /
2671 RW_MGR_MEM_VIRTUAL_GROUPS_PER_WRITE_DQS);
2672 rw_mgr_mem_calibrate_write_test_issue(write_group *
2673 RW_MGR_MEM_VIRTUAL_GROUPS_PER_WRITE_DQS+vg,
2676 base_rw_mgr = readl(addr_rw_mgr);
2677 tmp_bit_chk = tmp_bit_chk | (correct_mask_vg & ~(base_rw_mgr));
2681 *bit_chk &= tmp_bit_chk;
2685 set_rank_and_odt_mask(0, RW_MGR_ODT_MODE_OFF);
2686 debug_cond(DLEVEL == 2, "write_test(%u,%u,ALL) : %u == \
2687 %u => %lu", write_group, use_dm,
2688 *bit_chk, param->write_correct_mask,
2689 (long unsigned int)(*bit_chk ==
2690 param->write_correct_mask));
2691 return *bit_chk == param->write_correct_mask;
2693 set_rank_and_odt_mask(0, RW_MGR_ODT_MODE_OFF);
2694 debug_cond(DLEVEL == 2, "write_test(%u,%u,ONE) : %u != ",
2695 write_group, use_dm, *bit_chk);
2696 debug_cond(DLEVEL == 2, "%lu" " => %lu", (long unsigned int)0,
2697 (long unsigned int)(*bit_chk != 0));
2698 return *bit_chk != 0x00;
2703 * center all windows. do per-bit-deskew to possibly increase size of
2706 static uint32_t rw_mgr_mem_calibrate_writes_center(uint32_t rank_bgn,
2707 uint32_t write_group, uint32_t test_bgn)
2709 uint32_t i, p, min_index;
2712 * Store these as signed since there are comparisons with
2716 uint32_t sticky_bit_chk;
2717 int32_t left_edge[RW_MGR_MEM_DQ_PER_WRITE_DQS];
2718 int32_t right_edge[RW_MGR_MEM_DQ_PER_WRITE_DQS];
2720 int32_t mid_min, orig_mid_min;
2721 int32_t new_dqs, start_dqs, shift_dq;
2722 int32_t dq_margin, dqs_margin, dm_margin;
2724 uint32_t temp_dq_out1_delay;
2727 debug("%s:%d %u %u", __func__, __LINE__, write_group, test_bgn);
2731 addr = SDR_PHYGRP_SCCGRP_ADDRESS | SCC_MGR_IO_OUT1_DELAY_OFFSET;
2732 start_dqs = readl(addr +
2733 (RW_MGR_MEM_DQ_PER_WRITE_DQS << 2));
2735 /* per-bit deskew */
2738 * set the left and right edge of each bit to an illegal value
2739 * use (IO_IO_OUT1_DELAY_MAX + 1) as an illegal value.
2742 for (i = 0; i < RW_MGR_MEM_DQ_PER_WRITE_DQS; i++) {
2743 left_edge[i] = IO_IO_OUT1_DELAY_MAX + 1;
2744 right_edge[i] = IO_IO_OUT1_DELAY_MAX + 1;
2747 /* Search for the left edge of the window for each bit */
2748 for (d = 0; d <= IO_IO_OUT1_DELAY_MAX; d++) {
2749 scc_mgr_apply_group_dq_out1_delay(write_group, d);
2751 writel(0, &sdr_scc_mgr->update);
2754 * Stop searching when the read test doesn't pass AND when
2755 * we've seen a passing read on every bit.
2757 stop = !rw_mgr_mem_calibrate_write_test(rank_bgn, write_group,
2758 0, PASS_ONE_BIT, &bit_chk, 0);
2759 sticky_bit_chk = sticky_bit_chk | bit_chk;
2760 stop = stop && (sticky_bit_chk == param->write_correct_mask);
2761 debug_cond(DLEVEL == 2, "write_center(left): dtap=%d => %u \
2762 == %u && %u [bit_chk= %u ]\n",
2763 d, sticky_bit_chk, param->write_correct_mask,
2769 for (i = 0; i < RW_MGR_MEM_DQ_PER_WRITE_DQS; i++) {
2772 * Remember a passing test as the
2778 * If a left edge has not been seen
2779 * yet, then a future passing test will
2780 * mark this edge as the right edge.
2783 IO_IO_OUT1_DELAY_MAX + 1) {
2784 right_edge[i] = -(d + 1);
2787 debug_cond(DLEVEL == 2, "write_center[l,d=%d):", d);
2788 debug_cond(DLEVEL == 2, "bit_chk_test=%d left_edge[%u]: %d",
2789 (int)(bit_chk & 1), i, left_edge[i]);
2790 debug_cond(DLEVEL == 2, "right_edge[%u]: %d\n", i,
2792 bit_chk = bit_chk >> 1;
2797 /* Reset DQ delay chains to 0 */
2798 scc_mgr_apply_group_dq_out1_delay(0);
2800 for (i = RW_MGR_MEM_DQ_PER_WRITE_DQS - 1;; i--) {
2801 debug_cond(DLEVEL == 2, "%s:%d write_center: left_edge[%u]: \
2802 %d right_edge[%u]: %d\n", __func__, __LINE__,
2803 i, left_edge[i], i, right_edge[i]);
2806 * Check for cases where we haven't found the left edge,
2807 * which makes our assignment of the the right edge invalid.
2808 * Reset it to the illegal value.
2810 if ((left_edge[i] == IO_IO_OUT1_DELAY_MAX + 1) &&
2811 (right_edge[i] != IO_IO_OUT1_DELAY_MAX + 1)) {
2812 right_edge[i] = IO_IO_OUT1_DELAY_MAX + 1;
2813 debug_cond(DLEVEL == 2, "%s:%d write_center: reset \
2814 right_edge[%u]: %d\n", __func__, __LINE__,
2819 * Reset sticky bit (except for bits where we have
2820 * seen the left edge).
2822 sticky_bit_chk = sticky_bit_chk << 1;
2823 if ((left_edge[i] != IO_IO_OUT1_DELAY_MAX + 1))
2824 sticky_bit_chk = sticky_bit_chk | 1;
2830 /* Search for the right edge of the window for each bit */
2831 for (d = 0; d <= IO_IO_OUT1_DELAY_MAX - start_dqs; d++) {
2832 scc_mgr_apply_group_dqs_io_and_oct_out1(write_group,
2835 writel(0, &sdr_scc_mgr->update);
2838 * Stop searching when the read test doesn't pass AND when
2839 * we've seen a passing read on every bit.
2841 stop = !rw_mgr_mem_calibrate_write_test(rank_bgn, write_group,
2842 0, PASS_ONE_BIT, &bit_chk, 0);
2844 sticky_bit_chk = sticky_bit_chk | bit_chk;
2845 stop = stop && (sticky_bit_chk == param->write_correct_mask);
2847 debug_cond(DLEVEL == 2, "write_center (right): dtap=%u => %u == \
2848 %u && %u\n", d, sticky_bit_chk,
2849 param->write_correct_mask, stop);
2853 for (i = 0; i < RW_MGR_MEM_DQ_PER_WRITE_DQS;
2855 /* d = 0 failed, but it passed when
2856 testing the left edge, so it must be
2857 marginal, set it to -1 */
2858 if (right_edge[i] ==
2859 IO_IO_OUT1_DELAY_MAX + 1 &&
2861 IO_IO_OUT1_DELAY_MAX + 1) {
2868 for (i = 0; i < RW_MGR_MEM_DQ_PER_WRITE_DQS; i++) {
2871 * Remember a passing test as
2878 * If a right edge has not
2879 * been seen yet, then a future
2880 * passing test will mark this
2881 * edge as the left edge.
2883 if (right_edge[i] ==
2884 IO_IO_OUT1_DELAY_MAX + 1)
2885 left_edge[i] = -(d + 1);
2888 * d = 0 failed, but it passed
2889 * when testing the left edge,
2890 * so it must be marginal, set
2893 if (right_edge[i] ==
2894 IO_IO_OUT1_DELAY_MAX + 1 &&
2896 IO_IO_OUT1_DELAY_MAX + 1)
2899 * If a right edge has not been
2900 * seen yet, then a future
2901 * passing test will mark this
2902 * edge as the left edge.
2904 else if (right_edge[i] ==
2905 IO_IO_OUT1_DELAY_MAX +
2907 left_edge[i] = -(d + 1);
2910 debug_cond(DLEVEL == 2, "write_center[r,d=%d):", d);
2911 debug_cond(DLEVEL == 2, "bit_chk_test=%d left_edge[%u]: %d",
2912 (int)(bit_chk & 1), i, left_edge[i]);
2913 debug_cond(DLEVEL == 2, "right_edge[%u]: %d\n", i,
2915 bit_chk = bit_chk >> 1;
2920 /* Check that all bits have a window */
2921 for (i = 0; i < RW_MGR_MEM_DQ_PER_WRITE_DQS; i++) {
2922 debug_cond(DLEVEL == 2, "%s:%d write_center: left_edge[%u]: \
2923 %d right_edge[%u]: %d", __func__, __LINE__,
2924 i, left_edge[i], i, right_edge[i]);
2925 if ((left_edge[i] == IO_IO_OUT1_DELAY_MAX + 1) ||
2926 (right_edge[i] == IO_IO_OUT1_DELAY_MAX + 1)) {
2927 set_failing_group_stage(test_bgn + i,
2929 CAL_SUBSTAGE_WRITES_CENTER);
2934 /* Find middle of window for each DQ bit */
2935 mid_min = left_edge[0] - right_edge[0];
2937 for (i = 1; i < RW_MGR_MEM_DQ_PER_WRITE_DQS; i++) {
2938 mid = left_edge[i] - right_edge[i];
2939 if (mid < mid_min) {
2946 * -mid_min/2 represents the amount that we need to move DQS.
2947 * If mid_min is odd and positive we'll need to add one to
2948 * make sure the rounding in further calculations is correct
2949 * (always bias to the right), so just add 1 for all positive values.
2953 mid_min = mid_min / 2;
2954 debug_cond(DLEVEL == 1, "%s:%d write_center: mid_min=%d\n", __func__,
2957 /* Determine the amount we can change DQS (which is -mid_min) */
2958 orig_mid_min = mid_min;
2959 new_dqs = start_dqs;
2961 debug_cond(DLEVEL == 1, "%s:%d write_center: start_dqs=%d new_dqs=%d \
2962 mid_min=%d\n", __func__, __LINE__, start_dqs, new_dqs, mid_min);
2963 /* Initialize data for export structures */
2964 dqs_margin = IO_IO_OUT1_DELAY_MAX + 1;
2965 dq_margin = IO_IO_OUT1_DELAY_MAX + 1;
2967 /* add delay to bring centre of all DQ windows to the same "level" */
2968 for (i = 0, p = test_bgn; i < RW_MGR_MEM_DQ_PER_WRITE_DQS; i++, p++) {
2969 /* Use values before divide by 2 to reduce round off error */
2970 shift_dq = (left_edge[i] - right_edge[i] -
2971 (left_edge[min_index] - right_edge[min_index]))/2 +
2972 (orig_mid_min - mid_min);
2974 debug_cond(DLEVEL == 2, "%s:%d write_center: before: shift_dq \
2975 [%u]=%d\n", __func__, __LINE__, i, shift_dq);
2977 addr = SDR_PHYGRP_SCCGRP_ADDRESS | SCC_MGR_IO_OUT1_DELAY_OFFSET;
2978 temp_dq_out1_delay = readl(addr + (i << 2));
2979 if (shift_dq + (int32_t)temp_dq_out1_delay >
2980 (int32_t)IO_IO_OUT1_DELAY_MAX) {
2981 shift_dq = (int32_t)IO_IO_OUT1_DELAY_MAX - temp_dq_out1_delay;
2982 } else if (shift_dq + (int32_t)temp_dq_out1_delay < 0) {
2983 shift_dq = -(int32_t)temp_dq_out1_delay;
2985 debug_cond(DLEVEL == 2, "write_center: after: shift_dq[%u]=%d\n",
2987 scc_mgr_set_dq_out1_delay(i, temp_dq_out1_delay + shift_dq);
2990 debug_cond(DLEVEL == 2, "write_center: margin[%u]=[%d,%d]\n", i,
2991 left_edge[i] - shift_dq + (-mid_min),
2992 right_edge[i] + shift_dq - (-mid_min));
2993 /* To determine values for export structures */
2994 if (left_edge[i] - shift_dq + (-mid_min) < dq_margin)
2995 dq_margin = left_edge[i] - shift_dq + (-mid_min);
2997 if (right_edge[i] + shift_dq - (-mid_min) < dqs_margin)
2998 dqs_margin = right_edge[i] + shift_dq - (-mid_min);
3002 scc_mgr_apply_group_dqs_io_and_oct_out1(write_group, new_dqs);
3003 writel(0, &sdr_scc_mgr->update);
3006 debug_cond(DLEVEL == 2, "%s:%d write_center: DM\n", __func__, __LINE__);
3009 * set the left and right edge of each bit to an illegal value,
3010 * use (IO_IO_OUT1_DELAY_MAX + 1) as an illegal value,
3012 left_edge[0] = IO_IO_OUT1_DELAY_MAX + 1;
3013 right_edge[0] = IO_IO_OUT1_DELAY_MAX + 1;
3014 int32_t bgn_curr = IO_IO_OUT1_DELAY_MAX + 1;
3015 int32_t end_curr = IO_IO_OUT1_DELAY_MAX + 1;
3016 int32_t bgn_best = IO_IO_OUT1_DELAY_MAX + 1;
3017 int32_t end_best = IO_IO_OUT1_DELAY_MAX + 1;
3018 int32_t win_best = 0;
3020 /* Search for the/part of the window with DM shift */
3021 for (d = IO_IO_OUT1_DELAY_MAX; d >= 0; d -= DELTA_D) {
3022 scc_mgr_apply_group_dm_out1_delay(d);
3023 writel(0, &sdr_scc_mgr->update);
3025 if (rw_mgr_mem_calibrate_write_test(rank_bgn, write_group, 1,
3026 PASS_ALL_BITS, &bit_chk,
3028 /* USE Set current end of the window */
3031 * If a starting edge of our window has not been seen
3032 * this is our current start of the DM window.
3034 if (bgn_curr == IO_IO_OUT1_DELAY_MAX + 1)
3038 * If current window is bigger than best seen.
3039 * Set best seen to be current window.
3041 if ((end_curr-bgn_curr+1) > win_best) {
3042 win_best = end_curr-bgn_curr+1;
3043 bgn_best = bgn_curr;
3044 end_best = end_curr;
3047 /* We just saw a failing test. Reset temp edge */
3048 bgn_curr = IO_IO_OUT1_DELAY_MAX + 1;
3049 end_curr = IO_IO_OUT1_DELAY_MAX + 1;
3054 /* Reset DM delay chains to 0 */
3055 scc_mgr_apply_group_dm_out1_delay(0);
3058 * Check to see if the current window nudges up aganist 0 delay.
3059 * If so we need to continue the search by shifting DQS otherwise DQS
3060 * search begins as a new search. */
3061 if (end_curr != 0) {
3062 bgn_curr = IO_IO_OUT1_DELAY_MAX + 1;
3063 end_curr = IO_IO_OUT1_DELAY_MAX + 1;
3066 /* Search for the/part of the window with DQS shifts */
3067 for (d = 0; d <= IO_IO_OUT1_DELAY_MAX - new_dqs; d += DELTA_D) {
3069 * Note: This only shifts DQS, so are we limiting ourselve to
3070 * width of DQ unnecessarily.
3072 scc_mgr_apply_group_dqs_io_and_oct_out1(write_group,
3075 writel(0, &sdr_scc_mgr->update);
3076 if (rw_mgr_mem_calibrate_write_test(rank_bgn, write_group, 1,
3077 PASS_ALL_BITS, &bit_chk,
3079 /* USE Set current end of the window */
3082 * If a beginning edge of our window has not been seen
3083 * this is our current begin of the DM window.
3085 if (bgn_curr == IO_IO_OUT1_DELAY_MAX + 1)
3089 * If current window is bigger than best seen. Set best
3090 * seen to be current window.
3092 if ((end_curr-bgn_curr+1) > win_best) {
3093 win_best = end_curr-bgn_curr+1;
3094 bgn_best = bgn_curr;
3095 end_best = end_curr;
3098 /* We just saw a failing test. Reset temp edge */
3099 bgn_curr = IO_IO_OUT1_DELAY_MAX + 1;
3100 end_curr = IO_IO_OUT1_DELAY_MAX + 1;
3102 /* Early exit optimization: if ther remaining delay
3103 chain space is less than already seen largest window
3106 (IO_IO_OUT1_DELAY_MAX - new_dqs - d)) {
3112 /* assign left and right edge for cal and reporting; */
3113 left_edge[0] = -1*bgn_best;
3114 right_edge[0] = end_best;
3116 debug_cond(DLEVEL == 2, "%s:%d dm_calib: left=%d right=%d\n", __func__,
3117 __LINE__, left_edge[0], right_edge[0]);
3119 /* Move DQS (back to orig) */
3120 scc_mgr_apply_group_dqs_io_and_oct_out1(write_group, new_dqs);
3124 /* Find middle of window for the DM bit */
3125 mid = (left_edge[0] - right_edge[0]) / 2;
3127 /* only move right, since we are not moving DQS/DQ */
3131 /* dm_marign should fail if we never find a window */
3135 dm_margin = left_edge[0] - mid;
3137 scc_mgr_apply_group_dm_out1_delay(mid);
3138 writel(0, &sdr_scc_mgr->update);
3140 debug_cond(DLEVEL == 2, "%s:%d dm_calib: left=%d right=%d mid=%d \
3141 dm_margin=%d\n", __func__, __LINE__, left_edge[0],
3142 right_edge[0], mid, dm_margin);
3144 gbl->fom_out += dq_margin + dqs_margin;
3146 debug_cond(DLEVEL == 2, "%s:%d write_center: dq_margin=%d \
3147 dqs_margin=%d dm_margin=%d\n", __func__, __LINE__,
3148 dq_margin, dqs_margin, dm_margin);
3151 * Do not remove this line as it makes sure all of our
3152 * decisions have been applied.
3154 writel(0, &sdr_scc_mgr->update);
3155 return (dq_margin >= 0) && (dqs_margin >= 0) && (dm_margin >= 0);
3158 /* calibrate the write operations */
3159 static uint32_t rw_mgr_mem_calibrate_writes(uint32_t rank_bgn, uint32_t g,
3162 /* update info for sims */
3163 debug("%s:%d %u %u\n", __func__, __LINE__, g, test_bgn);
3165 reg_file_set_stage(CAL_STAGE_WRITES);
3166 reg_file_set_sub_stage(CAL_SUBSTAGE_WRITES_CENTER);
3168 reg_file_set_group(g);
3170 if (!rw_mgr_mem_calibrate_writes_center(rank_bgn, g, test_bgn)) {
3171 set_failing_group_stage(g, CAL_STAGE_WRITES,
3172 CAL_SUBSTAGE_WRITES_CENTER);
3180 * mem_precharge_and_activate() - Precharge all banks and activate
3182 * Precharge all banks and activate row 0 in bank "000..." and bank "111...".
3184 static void mem_precharge_and_activate(void)
3188 for (r = 0; r < RW_MGR_MEM_NUMBER_OF_RANKS; r++) {
3189 /* Test if the rank should be skipped. */
3190 if (param->skip_ranks[r])
3194 set_rank_and_odt_mask(r, RW_MGR_ODT_MODE_OFF);
3196 /* Precharge all banks. */
3197 writel(RW_MGR_PRECHARGE_ALL, SDR_PHYGRP_RWMGRGRP_ADDRESS |
3198 RW_MGR_RUN_SINGLE_GROUP_OFFSET);
3200 writel(0x0F, &sdr_rw_load_mgr_regs->load_cntr0);
3201 writel(RW_MGR_ACTIVATE_0_AND_1_WAIT1,
3202 &sdr_rw_load_jump_mgr_regs->load_jump_add0);
3204 writel(0x0F, &sdr_rw_load_mgr_regs->load_cntr1);
3205 writel(RW_MGR_ACTIVATE_0_AND_1_WAIT2,
3206 &sdr_rw_load_jump_mgr_regs->load_jump_add1);
3208 /* Activate rows. */
3209 writel(RW_MGR_ACTIVATE_0_AND_1, SDR_PHYGRP_RWMGRGRP_ADDRESS |
3210 RW_MGR_RUN_SINGLE_GROUP_OFFSET);
3215 * mem_init_latency() - Configure memory RLAT and WLAT settings
3217 * Configure memory RLAT and WLAT parameters.
3219 static void mem_init_latency(void)
3222 * For AV/CV, LFIFO is hardened and always runs at full rate
3223 * so max latency in AFI clocks, used here, is correspondingly
3226 const u32 max_latency = (1 << MAX_LATENCY_COUNT_WIDTH) - 1;
3229 debug("%s:%d\n", __func__, __LINE__);
3232 * Read in write latency.
3233 * WL for Hard PHY does not include additive latency.
3235 wlat = readl(&data_mgr->t_wl_add);
3236 wlat += readl(&data_mgr->mem_t_add);
3238 gbl->rw_wl_nop_cycles = wlat - 1;
3240 /* Read in readl latency. */
3241 rlat = readl(&data_mgr->t_rl_add);
3243 /* Set a pretty high read latency initially. */
3244 gbl->curr_read_lat = rlat + 16;
3245 if (gbl->curr_read_lat > max_latency)
3246 gbl->curr_read_lat = max_latency;
3248 writel(gbl->curr_read_lat, &phy_mgr_cfg->phy_rlat);
3250 /* Advertise write latency. */
3251 writel(wlat, &phy_mgr_cfg->afi_wlat);
3255 * @mem_skip_calibrate() - Set VFIFO and LFIFO to instant-on settings
3257 * Set VFIFO and LFIFO to instant-on settings in skip calibration mode.
3259 static void mem_skip_calibrate(void)
3261 uint32_t vfifo_offset;
3264 debug("%s:%d\n", __func__, __LINE__);
3265 /* Need to update every shadow register set used by the interface */
3266 for (r = 0; r < RW_MGR_MEM_NUMBER_OF_RANKS;
3267 r += NUM_RANKS_PER_SHADOW_REG) {
3269 * Set output phase alignment settings appropriate for
3272 for (i = 0; i < RW_MGR_MEM_IF_READ_DQS_WIDTH; i++) {
3273 scc_mgr_set_dqs_en_phase(i, 0);
3274 #if IO_DLL_CHAIN_LENGTH == 6
3275 scc_mgr_set_dqdqs_output_phase(i, 6);
3277 scc_mgr_set_dqdqs_output_phase(i, 7);
3282 * Write data arrives to the I/O two cycles before write
3283 * latency is reached (720 deg).
3284 * -> due to bit-slip in a/c bus
3285 * -> to allow board skew where dqs is longer than ck
3286 * -> how often can this happen!?
3287 * -> can claim back some ptaps for high freq
3288 * support if we can relax this, but i digress...
3290 * The write_clk leads mem_ck by 90 deg
3291 * The minimum ptap of the OPA is 180 deg
3292 * Each ptap has (360 / IO_DLL_CHAIN_LENGH) deg of delay
3293 * The write_clk is always delayed by 2 ptaps
3295 * Hence, to make DQS aligned to CK, we need to delay
3297 * (720 - 90 - 180 - 2 * (360 / IO_DLL_CHAIN_LENGTH))
3299 * Dividing the above by (360 / IO_DLL_CHAIN_LENGTH)
3300 * gives us the number of ptaps, which simplies to:
3302 * (1.25 * IO_DLL_CHAIN_LENGTH - 2)
3304 scc_mgr_set_dqdqs_output_phase(i,
3305 1.25 * IO_DLL_CHAIN_LENGTH - 2);
3307 writel(0xff, &sdr_scc_mgr->dqs_ena);
3308 writel(0xff, &sdr_scc_mgr->dqs_io_ena);
3310 for (i = 0; i < RW_MGR_MEM_IF_WRITE_DQS_WIDTH; i++) {
3311 writel(i, SDR_PHYGRP_SCCGRP_ADDRESS |
3312 SCC_MGR_GROUP_COUNTER_OFFSET);
3314 writel(0xff, &sdr_scc_mgr->dq_ena);
3315 writel(0xff, &sdr_scc_mgr->dm_ena);
3316 writel(0, &sdr_scc_mgr->update);
3319 /* Compensate for simulation model behaviour */
3320 for (i = 0; i < RW_MGR_MEM_IF_READ_DQS_WIDTH; i++) {
3321 scc_mgr_set_dqs_bus_in_delay(i, 10);
3322 scc_mgr_load_dqs(i);
3324 writel(0, &sdr_scc_mgr->update);
3327 * ArriaV has hard FIFOs that can only be initialized by incrementing
3330 vfifo_offset = CALIB_VFIFO_OFFSET;
3331 for (j = 0; j < vfifo_offset; j++)
3332 writel(0xff, &phy_mgr_cmd->inc_vfifo_hard_phy);
3333 writel(0, &phy_mgr_cmd->fifo_reset);
3336 * For Arria V and Cyclone V with hard LFIFO, we get the skip-cal
3337 * setting from generation-time constant.
3339 gbl->curr_read_lat = CALIB_LFIFO_OFFSET;
3340 writel(gbl->curr_read_lat, &phy_mgr_cfg->phy_rlat);
3344 * mem_calibrate() - Memory calibration entry point.
3346 * Perform memory calibration.
3348 static uint32_t mem_calibrate(void)
3351 uint32_t rank_bgn, sr;
3352 uint32_t write_group, write_test_bgn;
3353 uint32_t read_group, read_test_bgn;
3354 uint32_t run_groups, current_run;
3355 uint32_t failing_groups = 0;
3356 uint32_t group_failed = 0;
3358 const u32 rwdqs_ratio = RW_MGR_MEM_IF_READ_DQS_WIDTH /
3359 RW_MGR_MEM_IF_WRITE_DQS_WIDTH;
3361 debug("%s:%d\n", __func__, __LINE__);
3363 /* Initialize the data settings */
3364 gbl->error_substage = CAL_SUBSTAGE_NIL;
3365 gbl->error_stage = CAL_STAGE_NIL;
3366 gbl->error_group = 0xff;
3370 /* Initialize WLAT and RLAT. */
3373 /* Initialize bit slips. */
3374 mem_precharge_and_activate();
3376 for (i = 0; i < RW_MGR_MEM_IF_READ_DQS_WIDTH; i++) {
3377 writel(i, SDR_PHYGRP_SCCGRP_ADDRESS |
3378 SCC_MGR_GROUP_COUNTER_OFFSET);
3379 /* Only needed once to set all groups, pins, DQ, DQS, DM. */
3381 scc_mgr_set_hhp_extras();
3383 scc_set_bypass_mode(i);
3386 /* Calibration is skipped. */
3387 if ((dyn_calib_steps & CALIB_SKIP_ALL) == CALIB_SKIP_ALL) {
3389 * Set VFIFO and LFIFO to instant-on settings in skip
3392 mem_skip_calibrate();
3395 * Do not remove this line as it makes sure all of our
3396 * decisions have been applied.
3398 writel(0, &sdr_scc_mgr->update);
3402 /* Calibration is not skipped. */
3403 for (i = 0; i < NUM_CALIB_REPEAT; i++) {
3405 * Zero all delay chain/phase settings for all
3406 * groups and all shadow register sets.
3410 run_groups = ~param->skip_groups;
3412 for (write_group = 0, write_test_bgn = 0; write_group
3413 < RW_MGR_MEM_IF_WRITE_DQS_WIDTH; write_group++,
3414 write_test_bgn += RW_MGR_MEM_DQ_PER_WRITE_DQS) {
3416 /* Initialize the group failure */
3419 current_run = run_groups & ((1 <<
3420 RW_MGR_NUM_DQS_PER_WRITE_GROUP) - 1);
3421 run_groups = run_groups >>
3422 RW_MGR_NUM_DQS_PER_WRITE_GROUP;
3424 if (current_run == 0)
3427 writel(write_group, SDR_PHYGRP_SCCGRP_ADDRESS |
3428 SCC_MGR_GROUP_COUNTER_OFFSET);
3429 scc_mgr_zero_group(write_group, 0);
3431 for (read_group = write_group * rwdqs_ratio,
3433 read_group < (write_group + 1) * rwdqs_ratio;
3435 read_test_bgn += RW_MGR_MEM_DQ_PER_READ_DQS) {
3436 if (STATIC_CALIB_STEPS & CALIB_SKIP_VFIFO)
3439 /* Calibrate the VFIFO */
3440 if (rw_mgr_mem_calibrate_vfifo(read_group,
3444 if (!(gbl->phy_debug_mode_flags & PHY_DEBUG_SWEEP_ALL_GROUPS))
3447 /* The group failed, we're done. */
3451 /* Calibrate the output side */
3452 for (rank_bgn = 0, sr = 0;
3453 rank_bgn < RW_MGR_MEM_NUMBER_OF_RANKS;
3454 rank_bgn += NUM_RANKS_PER_SHADOW_REG, sr++) {
3455 if (STATIC_CALIB_STEPS & CALIB_SKIP_WRITES)
3458 /* Not needed in quick mode! */
3459 if (STATIC_CALIB_STEPS & CALIB_SKIP_DELAY_SWEEPS)
3463 * Determine if this set of ranks
3464 * should be skipped entirely.
3466 if (param->skip_shadow_regs[sr])
3469 /* Calibrate WRITEs */
3470 if (rw_mgr_mem_calibrate_writes(rank_bgn,
3471 write_group, write_test_bgn))
3475 if (!(gbl->phy_debug_mode_flags & PHY_DEBUG_SWEEP_ALL_GROUPS))
3479 /* Some group failed, we're done. */
3483 for (read_group = write_group * rwdqs_ratio,
3485 read_group < (write_group + 1) * rwdqs_ratio;
3487 read_test_bgn += RW_MGR_MEM_DQ_PER_READ_DQS) {
3488 if (STATIC_CALIB_STEPS & CALIB_SKIP_WRITES)
3491 if (rw_mgr_mem_calibrate_vfifo_end(read_group,
3495 if (!(gbl->phy_debug_mode_flags & PHY_DEBUG_SWEEP_ALL_GROUPS))
3498 /* The group failed, we're done. */
3502 /* No group failed, continue as usual. */
3505 grp_failed: /* A group failed, increment the counter. */
3510 * USER If there are any failing groups then report
3513 if (failing_groups != 0)
3516 if (STATIC_CALIB_STEPS & CALIB_SKIP_LFIFO)
3520 * If we're skipping groups as part of debug,
3521 * don't calibrate LFIFO.
3523 if (param->skip_groups != 0)
3526 /* Calibrate the LFIFO */
3527 if (!rw_mgr_mem_calibrate_lfifo())
3532 * Do not remove this line as it makes sure all of our decisions
3533 * have been applied.
3535 writel(0, &sdr_scc_mgr->update);
3540 * run_mem_calibrate() - Perform memory calibration
3542 * This function triggers the entire memory calibration procedure.
3544 static int run_mem_calibrate(void)
3548 debug("%s:%d\n", __func__, __LINE__);
3550 /* Reset pass/fail status shown on afi_cal_success/fail */
3551 writel(PHY_MGR_CAL_RESET, &phy_mgr_cfg->cal_status);
3553 /* Stop tracking manager. */
3554 clrbits_le32(&sdr_ctrl->ctrl_cfg, 1 << 22);
3556 phy_mgr_initialize();
3557 rw_mgr_mem_initialize();
3559 /* Perform the actual memory calibration. */
3560 pass = mem_calibrate();
3562 mem_precharge_and_activate();
3563 writel(0, &phy_mgr_cmd->fifo_reset);
3566 rw_mgr_mem_handoff();
3568 * In Hard PHY this is a 2-bit control:
3570 * 1: DDIO Mux Select
3572 writel(0x2, &phy_mgr_cfg->mux_sel);
3574 /* Start tracking manager. */
3575 setbits_le32(&sdr_ctrl->ctrl_cfg, 1 << 22);
3581 * debug_mem_calibrate() - Report result of memory calibration
3582 * @pass: Value indicating whether calibration passed or failed
3584 * This function reports the results of the memory calibration
3585 * and writes debug information into the register file.
3587 static void debug_mem_calibrate(int pass)
3589 uint32_t debug_info;
3592 printf("%s: CALIBRATION PASSED\n", __FILE__);
3597 if (gbl->fom_in > 0xff)
3600 if (gbl->fom_out > 0xff)
3601 gbl->fom_out = 0xff;
3603 /* Update the FOM in the register file */
3604 debug_info = gbl->fom_in;
3605 debug_info |= gbl->fom_out << 8;
3606 writel(debug_info, &sdr_reg_file->fom);
3608 writel(debug_info, &phy_mgr_cfg->cal_debug_info);
3609 writel(PHY_MGR_CAL_SUCCESS, &phy_mgr_cfg->cal_status);
3611 printf("%s: CALIBRATION FAILED\n", __FILE__);
3613 debug_info = gbl->error_stage;
3614 debug_info |= gbl->error_substage << 8;
3615 debug_info |= gbl->error_group << 16;
3617 writel(debug_info, &sdr_reg_file->failing_stage);
3618 writel(debug_info, &phy_mgr_cfg->cal_debug_info);
3619 writel(PHY_MGR_CAL_FAIL, &phy_mgr_cfg->cal_status);
3621 /* Update the failing group/stage in the register file */
3622 debug_info = gbl->error_stage;
3623 debug_info |= gbl->error_substage << 8;
3624 debug_info |= gbl->error_group << 16;
3625 writel(debug_info, &sdr_reg_file->failing_stage);
3628 printf("%s: Calibration complete\n", __FILE__);
3632 * hc_initialize_rom_data() - Initialize ROM data
3634 * Initialize ROM data.
3636 static void hc_initialize_rom_data(void)
3640 addr = SDR_PHYGRP_RWMGRGRP_ADDRESS | RW_MGR_INST_ROM_WRITE_OFFSET;
3641 for (i = 0; i < ARRAY_SIZE(inst_rom_init); i++)
3642 writel(inst_rom_init[i], addr + (i << 2));
3644 addr = SDR_PHYGRP_RWMGRGRP_ADDRESS | RW_MGR_AC_ROM_WRITE_OFFSET;
3645 for (i = 0; i < ARRAY_SIZE(ac_rom_init); i++)
3646 writel(ac_rom_init[i], addr + (i << 2));
3650 * initialize_reg_file() - Initialize SDR register file
3652 * Initialize SDR register file.
3654 static void initialize_reg_file(void)
3656 /* Initialize the register file with the correct data */
3657 writel(REG_FILE_INIT_SEQ_SIGNATURE, &sdr_reg_file->signature);
3658 writel(0, &sdr_reg_file->debug_data_addr);
3659 writel(0, &sdr_reg_file->cur_stage);
3660 writel(0, &sdr_reg_file->fom);
3661 writel(0, &sdr_reg_file->failing_stage);
3662 writel(0, &sdr_reg_file->debug1);
3663 writel(0, &sdr_reg_file->debug2);
3667 * initialize_hps_phy() - Initialize HPS PHY
3669 * Initialize HPS PHY.
3671 static void initialize_hps_phy(void)
3675 * Tracking also gets configured here because it's in the
3678 uint32_t trk_sample_count = 7500;
3679 uint32_t trk_long_idle_sample_count = (10 << 16) | 100;
3681 * Format is number of outer loops in the 16 MSB, sample
3686 reg |= SDR_CTRLGRP_PHYCTRL_PHYCTRL_0_ACDELAYEN_SET(2);
3687 reg |= SDR_CTRLGRP_PHYCTRL_PHYCTRL_0_DQDELAYEN_SET(1);
3688 reg |= SDR_CTRLGRP_PHYCTRL_PHYCTRL_0_DQSDELAYEN_SET(1);
3689 reg |= SDR_CTRLGRP_PHYCTRL_PHYCTRL_0_DQSLOGICDELAYEN_SET(1);
3690 reg |= SDR_CTRLGRP_PHYCTRL_PHYCTRL_0_RESETDELAYEN_SET(0);
3691 reg |= SDR_CTRLGRP_PHYCTRL_PHYCTRL_0_LPDDRDIS_SET(1);
3693 * This field selects the intrinsic latency to RDATA_EN/FULL path.
3694 * 00-bypass, 01- add 5 cycles, 10- add 10 cycles, 11- add 15 cycles.
3696 reg |= SDR_CTRLGRP_PHYCTRL_PHYCTRL_0_ADDLATSEL_SET(0);
3697 reg |= SDR_CTRLGRP_PHYCTRL_PHYCTRL_0_SAMPLECOUNT_19_0_SET(
3699 writel(reg, &sdr_ctrl->phy_ctrl0);
3702 reg |= SDR_CTRLGRP_PHYCTRL_PHYCTRL_1_SAMPLECOUNT_31_20_SET(
3704 SDR_CTRLGRP_PHYCTRL_PHYCTRL_0_SAMPLECOUNT_19_0_WIDTH);
3705 reg |= SDR_CTRLGRP_PHYCTRL_PHYCTRL_1_LONGIDLESAMPLECOUNT_19_0_SET(
3706 trk_long_idle_sample_count);
3707 writel(reg, &sdr_ctrl->phy_ctrl1);
3710 reg |= SDR_CTRLGRP_PHYCTRL_PHYCTRL_2_LONGIDLESAMPLECOUNT_31_20_SET(
3711 trk_long_idle_sample_count >>
3712 SDR_CTRLGRP_PHYCTRL_PHYCTRL_1_LONGIDLESAMPLECOUNT_19_0_WIDTH);
3713 writel(reg, &sdr_ctrl->phy_ctrl2);
3717 * initialize_tracking() - Initialize tracking
3719 * Initialize the register file with usable initial data.
3721 static void initialize_tracking(void)
3724 * Initialize the register file with the correct data.
3725 * Compute usable version of value in case we skip full
3726 * computation later.
3728 writel(DIV_ROUND_UP(IO_DELAY_PER_OPA_TAP, IO_DELAY_PER_DCHAIN_TAP) - 1,
3729 &sdr_reg_file->dtaps_per_ptap);
3731 /* trk_sample_count */
3732 writel(7500, &sdr_reg_file->trk_sample_count);
3734 /* longidle outer loop [15:0] */
3735 writel((10 << 16) | (100 << 0), &sdr_reg_file->trk_longidle);
3738 * longidle sample count [31:24]
3739 * trfc, worst case of 933Mhz 4Gb [23:16]
3740 * trcd, worst case [15:8]
3743 writel((243 << 24) | (14 << 16) | (10 << 8) | (4 << 0),
3744 &sdr_reg_file->delays);
3747 writel((RW_MGR_IDLE << 24) | (RW_MGR_ACTIVATE_1 << 16) |
3748 (RW_MGR_SGLE_READ << 8) | (RW_MGR_PRECHARGE_ALL << 0),
3749 &sdr_reg_file->trk_rw_mgr_addr);
3751 writel(RW_MGR_MEM_IF_READ_DQS_WIDTH,
3752 &sdr_reg_file->trk_read_dqs_width);
3755 writel((RW_MGR_REFRESH_ALL << 24) | (1000 << 0),
3756 &sdr_reg_file->trk_rfsh);
3759 int sdram_calibration_full(void)
3761 struct param_type my_param;
3762 struct gbl_type my_gbl;
3765 memset(&my_param, 0, sizeof(my_param));
3766 memset(&my_gbl, 0, sizeof(my_gbl));
3771 /* Set the calibration enabled by default */
3772 gbl->phy_debug_mode_flags |= PHY_DEBUG_ENABLE_CAL_RPT;
3774 * Only sweep all groups (regardless of fail state) by default
3775 * Set enabled read test by default.
3777 #if DISABLE_GUARANTEED_READ
3778 gbl->phy_debug_mode_flags |= PHY_DEBUG_DISABLE_GUARANTEED_READ;
3780 /* Initialize the register file */
3781 initialize_reg_file();
3783 /* Initialize any PHY CSR */
3784 initialize_hps_phy();
3786 scc_mgr_initialize();
3788 initialize_tracking();
3790 printf("%s: Preparing to start memory calibration\n", __FILE__);
3792 debug("%s:%d\n", __func__, __LINE__);
3793 debug_cond(DLEVEL == 1,
3794 "DDR3 FULL_RATE ranks=%u cs/dimm=%u dq/dqs=%u,%u vg/dqs=%u,%u ",
3795 RW_MGR_MEM_NUMBER_OF_RANKS, RW_MGR_MEM_NUMBER_OF_CS_PER_DIMM,
3796 RW_MGR_MEM_DQ_PER_READ_DQS, RW_MGR_MEM_DQ_PER_WRITE_DQS,
3797 RW_MGR_MEM_VIRTUAL_GROUPS_PER_READ_DQS,
3798 RW_MGR_MEM_VIRTUAL_GROUPS_PER_WRITE_DQS);
3799 debug_cond(DLEVEL == 1,
3800 "dqs=%u,%u dq=%u dm=%u ptap_delay=%u dtap_delay=%u ",
3801 RW_MGR_MEM_IF_READ_DQS_WIDTH, RW_MGR_MEM_IF_WRITE_DQS_WIDTH,
3802 RW_MGR_MEM_DATA_WIDTH, RW_MGR_MEM_DATA_MASK_WIDTH,
3803 IO_DELAY_PER_OPA_TAP, IO_DELAY_PER_DCHAIN_TAP);
3804 debug_cond(DLEVEL == 1, "dtap_dqsen_delay=%u, dll=%u",
3805 IO_DELAY_PER_DQS_EN_DCHAIN_TAP, IO_DLL_CHAIN_LENGTH);
3806 debug_cond(DLEVEL == 1, "max values: en_p=%u dqdqs_p=%u en_d=%u dqs_in_d=%u ",
3807 IO_DQS_EN_PHASE_MAX, IO_DQDQS_OUT_PHASE_MAX,
3808 IO_DQS_EN_DELAY_MAX, IO_DQS_IN_DELAY_MAX);
3809 debug_cond(DLEVEL == 1, "io_in_d=%u io_out1_d=%u io_out2_d=%u ",
3810 IO_IO_IN_DELAY_MAX, IO_IO_OUT1_DELAY_MAX,
3811 IO_IO_OUT2_DELAY_MAX);
3812 debug_cond(DLEVEL == 1, "dqs_in_reserve=%u dqs_out_reserve=%u\n",
3813 IO_DQS_IN_RESERVE, IO_DQS_OUT_RESERVE);
3815 hc_initialize_rom_data();
3817 /* update info for sims */
3818 reg_file_set_stage(CAL_STAGE_NIL);
3819 reg_file_set_group(0);
3822 * Load global needed for those actions that require
3823 * some dynamic calibration support.
3825 dyn_calib_steps = STATIC_CALIB_STEPS;
3827 * Load global to allow dynamic selection of delay loop settings
3828 * based on calibration mode.
3830 if (!(dyn_calib_steps & CALIB_SKIP_DELAY_LOOPS))
3831 skip_delay_mask = 0xff;
3833 skip_delay_mask = 0x0;
3835 pass = run_mem_calibrate();
3836 debug_mem_calibrate(pass);